netlink: Allow command handler to use l_netlink_cancel
PATCH 1/1 netlink: Allow command handler to use l_netlink_cancel
--- ell/netlink.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
ell/netlink.c +5 -3
@@ -162,12 +162,15 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) {const void *data = nlmsg;struct command *command; +uint32_t id; command =l_hashmap_remove (netlink->command_pending,L_UINT_TO_PTR (nlmsg->nlmsg_seq));if (!command)return ; + id = command->id; +if (!command->handler) goto done;
@@ -189,9 +192,8 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) } done: -l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (command->id)); - -destroy_command (command); +if (l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (id))) +destroy_command (command); }static void process_multi (struct l_netlink *netlink,struct nlmsghdr *nlmsg)
Signature · 1 line
2.27.0
Hi Andrew,
Quoted 4 lines · Andrew Zaborowski
--- ell/netlink.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
You may want to write something about why this would be wanted.
Quoted 21 lines
ell/netlink.c +3
-0@@ -162,12 +162,15 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) {const void *data = nlmsg;struct command *command; +uint32_t id; command =l_hashmap_remove (netlink->command_pending,L_UINT_TO_PTR (nlmsg->nlmsg_seq));if (!command)return ; + id = command->id; +if (!command->handler) goto done;
ell/netlink.c static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) · line 196
Quoted 8 lines
ell/netlink.c +2
-3@@ -189,9 +192,8 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) } done: -l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (command->id)); - -destroy_command (command); +if (l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (id))) +destroy_command (command);
Would it not be better to simply remove the command from the command_lookup hashmap prior to dispatch, and have l_netlink_cancel simply fail?
I do wonder why you'd cancel the command from the handler?
ell/netlink.c static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) · line 199
Quoted 4 lines
ell/netlink.c +0
-0}static void process_multi (struct l_netlink *netlink,struct nlmsghdr *nlmsg)
Regards,
-Denis
Quoted 7 lines · Denis Kenzior
Quoted 4 lines · Andrew Zaborowski
--- ell/netlink.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)You may want to write something about why this would be wanted.
Ok.
Quoted 24 lines
Quoted 24 lines
ell/netlink.c +3
-0@@ -162,12 +162,15 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) {const void *data = nlmsg;struct command *command; +uint32_t id; command =l_hashmap_remove (netlink->command_pending,L_UINT_TO_PTR (nlmsg->nlmsg_seq));if (!command)return ; + id = command->id; +if (!command->handler) goto done;@@ -189,9 +192,8 @@ static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) }
Quoted 9 lines
ell/netlink.c static void process_message(struct l_netlink *netlink, struct nlmsghdr *nlmsg) · line 196
Quoted 6 lines
ell/netlink.c +2
-3done: -l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (command->id)); - -destroy_command (command); +if (l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (id))) +destroy_command (command);Would it not be better to simply remove the command from the
command_lookuphashmap prior to dispatch, and havel_netlink_cancelsimply fail?
Yes, that would work.
Quoted 2 lines
I do wonder why you'd cancel the command from the handler?
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Best regards
Hi Andrew,
Quoted 7 lines
Quoted 1 line
I do wonder why you'd cancel the command from the handler?
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Sure, I buy the more robust argument, though I still wonder how one would even get into such a state on purpose. But the last sentence seems to be handwaving?
So did you somehow trigger this condition or just fixing ghost bugs?
Regards,
-Denis
Quoted 8 lines · Denis Kenzior
Quoted 5 lines
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Sure, I buy the more robust argument, though I still wonder how one would even get into such a state on purpose. But the last sentence seems to be handwaving?
Well, in the AP code we're actually careful to reset e.g. ap->rtnl_add_cmd to 0 at the beginning of the reply handler but if we were resetting it any later we'd be affected by this.
Quoted 1 line
So did you somehow trigger this condition or just fixing ghost bugs?
I was only checking the code to see if it was safe to cancel the dump command the way I did it in my IWD patchset. It seemed to be safe for dump commands but not for regular commands.
The normal sequence is this:
1. l_netlink_send call
2. handler callback
3. destroy callback
After 3. I'd consider it wrong to call l_netlink_cancel() with the original ID because you'd be risking cancelling an unrelated command (there's no guarantee that the ID wasn't re-used since), but until 3. I think it should be made safe.
Best regards
Hi Andrew
Quoted 7 lines · Andrew Zaborowski
Quoted 6 lines · Denis Kenzior
Quoted 5 lines
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Quoted 7 lines
Quoted 2 lines
Sure, I buy the more robust argument, though I still wonder how one would even get into such a state on purpose. But the last sentence seems to be handwaving?
Well, in the AP code we're actually careful to reset e.g.
ap->rtnl_add_cmdto 0 at the beginning of the reply handler but if we were resetting it any later we'd be affected by this.
Hmm... This is pretty typical pattern for when one doesn't provide a destroy callback. I'm not sure I understand what you're trying to convey here...?
Are you trying to stop callbacks mid-dump or ?
Quoted 5 lines
Quoted 1 line
So did you somehow trigger this condition or just fixing ghost bugs?
I was only checking the code to see if it was safe to cancel the dump command the way I did it in my IWD patchset. It seemed to be safe for dump commands but not for regular commands.
Yes, but *why* would you use l_netlink_cancel for the command when you're already in the handler for that exact same command?
Quoted 10 lines
The normal sequence is this:
1.l_netlink_sendcall
2. handler callback
3. destroy callbackAfter 3. I'd consider it wrong to call
l_netlink_cancel()with the original ID because you'd be risking cancelling an unrelated command (there's no guarantee that the ID wasn't re-used since), but until 3. I think it should be made safe.
I can't really imagine why anyone would invoke l_netlink_cancel in a callback. Not saying we shouldn't allow this, but the use case is still a mystery to me?
Also, we sort of cheat in many places and assume that only a single message will be sent for certain commands. Underneath, there's no real difference between a dump and a non-dump. In fact the kernel might (someday) choose to send multiple messages as a reply. The only way you know that you're done (for netlink/genl at least) is after the destroy callback has been called. The callback itself has no way of knowing whether it is a single-part or multi-part message...
Regards,
-Denis
Quoted 16 lines · Denis Kenzior
Quoted 7 lines · Andrew Zaborowski
Quoted 6 lines · Denis Kenzior
Quoted 5 lines
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Quoted 7 lines
Quoted 2 lines
Sure, I buy the more robust argument, though I still wonder how one would even get into such a state on purpose. But the last sentence seems to be handwaving?
Well, in the AP code we're actually careful to reset e.g.
ap->rtnl_add_cmdto 0 at the beginning of the reply handler but if we were resetting it any later we'd be affected by this.
Quoted 4 lines
Hmm... This is pretty typical pattern for when one doesn't provide a destroy callback. I'm not sure I understand what you're trying to convey here...?
Are you trying to stop callbacks mid-dump or ?
Yes, in the dump case the use case is pretty obvious -- once you've received the fragment you were interested in (in my case, I dump IPv4 addresses and I'm interested in one specific ifindex), if you can cancel the command you don't have to keep a flag that tells you to ignore further callbacks. And if you need to destroy the object (e.g. stop the AP) based on the data you received, you're also free to do this.
In the non-dump case I see the utility of this change mainly in the situation where, within the callback, you decide you want to tear-down the object (e.g. stop the AP), like I said already. If you're not allowed to cancel the command at that point, you basically need reference counting so that you can avoid crashing in your destroy handler. And in that case I'd say we should have at least a warning about this limitation somewhere.
Quoted 24 lines
Quoted 6 lines
Quoted 2 lines
So did you somehow trigger this condition or just fixing ghost bugs?
I was only checking the code to see if it was safe to cancel the dump command the way I did it in my IWD patchset. It seemed to be safe for dump commands but not for regular commands.
Yes, but *why* would you use
l_netlink_cancelfor the command when you're already in the handler for that exact same command?Quoted 10 lines
The normal sequence is this:
1.l_netlink_sendcall
2. handler callback
3. destroy callbackAfter 3. I'd consider it wrong to call
l_netlink_cancel()with the original ID because you'd be risking cancelling an unrelated command (there's no guarantee that the ID wasn't re-used since), but until 3. I think it should be made safe.I can't really imagine why anyone would invoke
l_netlink_cancelin a callback. Not saying we shouldn't allow this, but the use case is still a mystery to me?
Quoted 6 lines
Also, we sort of cheat in many places and assume that only a single message will be sent for certain commands. Underneath, there's no real difference between a dump and a non-dump. In fact the kernel might (someday) choose to send multiple messages as a reply. The only way you know that you're done (for netlink/genl at least) is after the destroy callback has been called. The callback itself has no way of knowing whether it is a single-part or multi-part message...
Oh right, the dump/non-dump distinction in my example is based on the current rtnl behavior.
Best regards
Hi Andrew,
Quoted 9 lines
Quoted 1 line
Are you trying to stop callbacks mid-dump or ?
Yes, in the dump case the use case is pretty obvious -- once you've received the fragment you were interested in (in my case, I dump IPv4 addresses and I'm interested in one specific ifindex), if you can cancel the command you don't have to keep a flag that tells you to ignore further callbacks. And if you need to destroy the object (e.g. stop the AP) based on the data you received, you're also free to do this.
Ok, gotcha. I have to ponder whether using l_foo_cancel is a good idea for this, or a more explicit API should be used. Also, we'd have to make this behavior consistent everywhere.
Quoted 8 lines
In the non-dump case I see the utility of this change mainly in the situation where, within the callback, you decide you want to tear-down the object (e.g. stop the AP), like I said already. If you're not allowed to cancel the command at that point, you basically need reference counting so that you can avoid crashing in your destroy handler. And in that case I'd say we should have at least a warning about this limitation somewhere.
You might be the first one to need this actually. There's always the l_idle_oneshot get-out-of-jail card ;)
Regards,
-Denis
Hi Denis,
Quoted 8 lines · Andrew Zaborowski
Quoted 7 lines · Denis Kenzior
Quoted 6 lines
done:
ell/netlink.c +2
-3-l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (command->id)); - -destroy_command (command); +if (l_hashmap_remove (netlink->command_lookup,L_UINT_TO_PTR (id))) +destroy_command (command);
Quoted 4 lines
Quoted 2 lines
Would it not be better to simply remove the command from the
command_lookuphashmap prior to dispatch, and havel_netlink_cancelsimply fail?Yes, that would work.
On a second thought, I think it's better the other way. The l_netlink_cancel() call can be used as a request to forget the callbacks. The user may be tearing down the object they passed as user_data for example. IIRC we did take that use case into consideration in some of the other APIs, what do you think?
Quoted 11 lines
Quoted 3 lines
I do wonder why you'd cancel the command from the handler?
In the IWD AP case I want to cancel a dump command which I think is already handled correctly, but if the API can be made more robust then I think it's an improvement. There can be complicated code paths where you destroy an object that owns the netlink command as a result of the command response.
Best regards
Hi Andrew,
Quoted 10 lines
Quoted 4 lines
Quoted 2 lines
Would it not be better to simply remove the command from the
command_lookuphashmap prior to dispatch, and havel_netlink_cancelsimply fail?Yes, that would work.
On a second thought, I think it's better the other way. The
l_netlink_cancel()call can be used as a request to forget the callbacks. The user may be tearing down the object they passed asuser_datafor example. IIRC we did take that use case into consideration in some of the other APIs, what do you think?
Which other way? Ahhh, I think I get it... You want to unset the destroy callback or prevent it from being called subsequently when inside the handler by issuing l_netlink_cancel?
I actually don't think any ell APIs work this way. l_dbus for example simply removes the pending request prior to dispatch (which is what I suggested earlier). Hence l_dbus_cancel wouldn't work on it inside the l_dbus_send callback.
Even APIs that use the watchlist pattern don't do it this way. You may be thinking of the scan API inside iwd, which did it for a different reason.
Regards,
-Denis