Re: [PATCH] bpf: devmap: allow for repeated redirect

[Florian Kauer]

On 9/11/24 11:25, Toke Høiland-Jørgensen wrote:
> Florian Kauer <florian.kauer@xxxxxxxxxxxxx> writes:
> 
>> Hi Toke,
>>
>> On 9/10/24 10:34, Toke Høiland-Jørgensen wrote:
>>> Florian Kauer <florian.kauer@xxxxxxxxxxxxx> writes:
>>>
>>>> Currently, returning XDP_REDIRECT from a xdp/devmap program
>>>> is considered as invalid action and an exception is traced.
>>>>
>>>> While it might seem counterintuitive to redirect in a xdp/devmap
>>>> program (why not just redirect to the correct interface in the
>>>> first program?), we faced several use cases where supporting
>>>> this would be very useful.
>>>>
>>>> Most importantly, they occur when the first redirect is used
>>>> with the BPF_F_BROADCAST flag. Using this together with xdp/devmap
>>>> programs, enables to perform different actions on clones of
>>>> the same incoming frame. In that case, it is often useful
>>>> to redirect either to a different CPU or device AFTER the cloning.
>>>>
>>>> For example:
>>>> - Replicate the frame (for redundancy according to IEEE 802.1CB FRER)
>>>>   and then use the second redirect with a cpumap to select
>>>>   the path-specific egress queue.
>>>>
>>>> - Also, one of the paths might need an encapsulation that
>>>>   exceeds the MTU. So a second redirect can be used back
>>>>   to the ingress interface to send an ICMP FRAG_NEEDED packet.
>>>>
>>>> - For OAM purposes, you might want to send one frame with
>>>>   OAM information back, while the original frame in passed forward.
>>>>
>>>> To enable these use cases, add the XDP_REDIRECT case to
>>>> dev_map_bpf_prog_run. Also, when this is called from inside
>>>> xdp_do_flush, the redirect might add further entries to the
>>>> flush lists that are currently processed. Therefore, loop inside
>>>> xdp_do_flush until no more additional redirects were added.
>>>>
>>>> Signed-off-by: Florian Kauer <florian.kauer@xxxxxxxxxxxxx>
>>>
>>> This is an interesting use case! However, your implementation makes it
>>> way to easy to end up in a situation that loops forever, so I think we
>>> should add some protection against that. Some details below:
>>>
>>>> ---
>>>>  include/linux/bpf.h        |  4 ++--
>>>>  include/trace/events/xdp.h | 10 ++++++----
>>>>  kernel/bpf/devmap.c        | 37 +++++++++++++++++++++++++++--------
>>>>  net/core/filter.c          | 48 +++++++++++++++++++++++++++-------------------
>>>>  4 files changed, 65 insertions(+), 34 deletions(-)
>>>>
>>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>>> index 3b94ec161e8c..1b57cbabf199 100644
>>>> --- a/include/linux/bpf.h
>>>> +++ b/include/linux/bpf.h
>>>> @@ -2498,7 +2498,7 @@ struct sk_buff;
>>>>  struct bpf_dtab_netdev;
>>>>  struct bpf_cpu_map_entry;
>>>>  
>>>> -void __dev_flush(struct list_head *flush_list);
>>>> +void __dev_flush(struct list_head *flush_list, int *redirects);
>>>>  int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
>>>>  		    struct net_device *dev_rx);
>>>>  int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
>>>> @@ -2740,7 +2740,7 @@ static inline struct bpf_token *bpf_token_get_from_fd(u32 ufd)
>>>>  	return ERR_PTR(-EOPNOTSUPP);
>>>>  }
>>>>  
>>>> -static inline void __dev_flush(struct list_head *flush_list)
>>>> +static inline void __dev_flush(struct list_head *flush_list, int *redirects)
>>>>  {
>>>>  }
>>>>  
>>>> diff --git a/include/trace/events/xdp.h b/include/trace/events/xdp.h
>>>> index a7e5452b5d21..fba2c457e727 100644
>>>> --- a/include/trace/events/xdp.h
>>>> +++ b/include/trace/events/xdp.h
>>>> @@ -269,9 +269,9 @@ TRACE_EVENT(xdp_devmap_xmit,
>>>>  
>>>>  	TP_PROTO(const struct net_device *from_dev,
>>>>  		 const struct net_device *to_dev,
>>>> -		 int sent, int drops, int err),
>>>> +		 int sent, int drops, int redirects, int err),
>>>>  
>>>> -	TP_ARGS(from_dev, to_dev, sent, drops, err),
>>>> +	TP_ARGS(from_dev, to_dev, sent, drops, redirects, err),
>>>>  
>>>>  	TP_STRUCT__entry(
>>>>  		__field(int, from_ifindex)
>>>> @@ -279,6 +279,7 @@ TRACE_EVENT(xdp_devmap_xmit,
>>>>  		__field(int, to_ifindex)
>>>>  		__field(int, drops)
>>>>  		__field(int, sent)
>>>> +		__field(int, redirects)
>>>>  		__field(int, err)
>>>>  	),
>>>>  
>>>> @@ -288,16 +289,17 @@ TRACE_EVENT(xdp_devmap_xmit,
>>>>  		__entry->to_ifindex	= to_dev->ifindex;
>>>>  		__entry->drops		= drops;
>>>>  		__entry->sent		= sent;
>>>> +		__entry->redirects	= redirects;
>>>>  		__entry->err		= err;
>>>>  	),
>>>>  
>>>>  	TP_printk("ndo_xdp_xmit"
>>>>  		  " from_ifindex=%d to_ifindex=%d action=%s"
>>>> -		  " sent=%d drops=%d"
>>>> +		  " sent=%d drops=%d redirects=%d"
>>>>  		  " err=%d",
>>>>  		  __entry->from_ifindex, __entry->to_ifindex,
>>>>  		  __print_symbolic(__entry->act, __XDP_ACT_SYM_TAB),
>>>> -		  __entry->sent, __entry->drops,
>>>> +		  __entry->sent, __entry->drops, __entry->redirects,
>>>>  		  __entry->err)
>>>>  );
>>>>  
>>>> diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
>>>> index 7878be18e9d2..89bdec49ea40 100644
>>>> --- a/kernel/bpf/devmap.c
>>>> +++ b/kernel/bpf/devmap.c
>>>> @@ -334,7 +334,8 @@ static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
>>>>  static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
>>>>  				struct xdp_frame **frames, int n,
>>>>  				struct net_device *tx_dev,
>>>> -				struct net_device *rx_dev)
>>>> +				struct net_device *rx_dev,
>>>> +				int *redirects)
>>>>  {
>>>>  	struct xdp_txq_info txq = { .dev = tx_dev };
>>>>  	struct xdp_rxq_info rxq = { .dev = rx_dev };
>>>> @@ -359,6 +360,13 @@ static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
>>>>  			else
>>>>  				frames[nframes++] = xdpf;
>>>>  			break;
>>>> +		case XDP_REDIRECT:
>>>> +			err = xdp_do_redirect(rx_dev, &xdp, xdp_prog);
>>>> +			if (unlikely(err))
>>>> +				xdp_return_frame_rx_napi(xdpf);
>>>> +			else
>>>> +				*redirects += 1;
>>>> +			break;
>>>
>>> It's a bit subtle, but dev_map_bpf_prog_run() also filters the list of
>>> frames in the queue in-place (the frames[nframes++] = xdpf; line above),
>>> which only works under the assumption that the array in bq->q is not
>>> modified while this loop is being run. But now you're adding a call in
>>> the middle that may result in the packet being put back on the same
>>> queue in the middle, which means that this assumption no longer holds.
>>>
>>> So you need to clear the bq->q queue first for this to work.
>>> Specifically, at the start of bq_xmit_all(), you'll need to first copy
>>> all the packet pointer onto an on-stack array, then run the rest of the
>>> function on that array. There's already an initial loop that goes
>>> through all the frames, so you can just do it there.
>>>
>>> So the loop at the start of bq_xmit_all() goes from the current:
>>>
>>> 	for (i = 0; i < cnt; i++) {
>>> 		struct xdp_frame *xdpf = bq->q[i];
>>>
>>> 		prefetch(xdpf);
>>> 	}
>>>
>>>
>>> to something like:
>>>
>>>         struct xdp_frame *frames[DEV_MAP_BULK_SIZE];
>>>
>>> 	for (i = 0; i < cnt; i++) {
>>> 		struct xdp_frame *xdpf = bq->q[i];
>>>
>>> 		prefetch(xdpf);
>>>                 frames[i] = xdpf;
>>> 	}
>>>
>>>         bq->count = 0; /* bq is now empty, use the 'frames' and 'cnt'
>>>                           stack variables for the rest of the function */
>>>
>>>
>>>
>>>>  		default:
>>>>  			bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
>>>>  			fallthrough;
>>>> @@ -373,7 +381,7 @@ static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
>>>>  	return nframes; /* sent frames count */
>>>>  }
>>>>  
>>>> -static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
>>>> +static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags, int *redirects)
>>>>  {
>>>>  	struct net_device *dev = bq->dev;
>>>>  	unsigned int cnt = bq->count;
>>>> @@ -390,8 +398,10 @@ static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
>>>>  		prefetch(xdpf);
>>>>  	}
>>>>  
>>>> +	int new_redirects = 0;
>>>>  	if (bq->xdp_prog) {
>>>> -		to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev, bq->dev_rx);
>>>> +		to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev, bq->dev_rx,
>>>> +				&new_redirects);
>>>>  		if (!to_send)
>>>>  			goto out;
>>>>  	}
>>>> @@ -413,19 +423,21 @@ static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
>>>>  
>>>>  out:
>>>>  	bq->count = 0;
>>>> -	trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
>>>> +	*redirects += new_redirects;
>>>> +	trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent - new_redirects,
>>>> +			new_redirects, err);
>>>>  }
>>>>  
>>>>  /* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
>>>>   * driver before returning from its napi->poll() routine. See the comment above
>>>>   * xdp_do_flush() in filter.c.
>>>>   */
>>>> -void __dev_flush(struct list_head *flush_list)
>>>> +void __dev_flush(struct list_head *flush_list, int *redirects)
>>>>  {
>>>>  	struct xdp_dev_bulk_queue *bq, *tmp;
>>>>  
>>>>  	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
>>>> -		bq_xmit_all(bq, XDP_XMIT_FLUSH);
>>>> +		bq_xmit_all(bq, XDP_XMIT_FLUSH, redirects);
>>>>  		bq->dev_rx = NULL;
>>>>  		bq->xdp_prog = NULL;
>>>>  		__list_del_clearprev(&bq->flush_node);
>>>> @@ -458,8 +470,17 @@ static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
>>>>  {
>>>>  	struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
>>>>  
>>>> -	if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
>>>> -		bq_xmit_all(bq, 0);
>>>> +	if (unlikely(bq->count == DEV_MAP_BULK_SIZE)) {
>>>> +		int redirects = 0;
>>>> +
>>>> +		bq_xmit_all(bq, 0, &redirects);
>>>> +
>>>> +		/* according to comment above xdp_do_flush() in
>>>> +		 * filter.c, xdp_do_flush is always called at
>>>> +		 * the end of the NAPI anyway, so no need to act
>>>> +		 * on the redirects here
>>>> +		 */
>>>
>>> While it's true that it will be called again in NAPI, the purpose of
>>> calling bq_xmit_all() here is to make room space for the packet on the
>>> bulk queue that we're about to enqueue, and if bq_xmit_all() can just
>>> put the packet back on the queue, there is no guarantee that this will
>>> succeed. So you will have to handle that case here.
>>>
>>> Since there's also a potential infinite recursion issue in the
>>> do_flush() functions below, I think it may be better to handle this
>>> looping issue inside bq_xmit_all().
>>>
>>> I.e., structure the code so that bq_xmit_all() guarantees that when it
>>> returns it has actually done its job; that is, that bq->q is empty.
>>>
>>> Given the above "move all frames out of bq->q at the start" change, this
>>> is not all that hard. Simply add a check after the out: label (in
>>> bq_xmit_all()) to check if bq->count is actually 0, and if it isn't,
>>> start over from the beginning of that function. This also makes it
>>> straight forward to add a recursion limit; after looping a set number of
>>> times (say, XMIT_RECURSION_LIMIT), simply turn XDP_REDIRECT into drops.
>>>
>>> There will need to be some additional protection against looping forever
>>> in __dev_flush(), to handle the case where a packet is looped between
>>> two interfaces. This one is a bit trickier, but a similar recursion
>>> counter could be used, I think.
>>
>>
>> Thanks a lot for the extensive support!
>> Regarding __dev_flush(), could the following already be sufficient?
>>
>> void __dev_flush(struct list_head *flush_list)
>> {
>>         struct xdp_dev_bulk_queue *bq, *tmp;
>>         int i,j;
>>
>>         for (i = 0; !list_empty(flush_list) && i < XMIT_RECURSION_LIMIT; i++) {
>>                 /* go through list in reverse so that new items
>>                  * added to the flush_list will only be handled
>>                  * in the next iteration of the outer loop
>>                  */
>>                 list_for_each_entry_safe_reverse(bq, tmp, flush_list, flush_node) {
>>                         bq_xmit_all(bq, XDP_XMIT_FLUSH);
>>                         bq->dev_rx = NULL;
>>                         bq->xdp_prog = NULL;
>>                         __list_del_clearprev(&bq->flush_node);
>>                 }
>>         }
>>
>>         if (i == XMIT_RECURSION_LIMIT) {
>>                 pr_warn("XDP recursion limit hit, expect packet loss!\n");
>>
>>                 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
>>                         for (j = 0; j < bq->count; j++)
>>                                 xdp_return_frame_rx_napi(bq->q[i]);
>>
>>                         bq->count = 0;
>>                         bq->dev_rx = NULL;
>>                         bq->xdp_prog = NULL;
>>                         __list_del_clearprev(&bq->flush_node);
>>                 }
>>         }
>> }
> 
> Yeah, this would work, I think (neat trick with iterating the list in
> reverse!), but instead of the extra loop in the end, I would suggest
> passing in an extra 'allow_redirect' parameter to bq_xmit_all(). Since
> you'll already have to handle the recursion limit inside that function,
> you can just reuse the same functionality by passing in the parameter in
> the last iteration of the first loop.
> 
> Also, definitely don't put an unconditional warn_on() in the fast path -
> that brings down the system really quickly if it's misconfigured :)
> 
> A warn_on_once() could work, but really I would suggest just reusing the
> _trace_xdp_redirect_map_err() tracepoint with a new error code (just has
> to be one we're not currently using and that vaguely resembles what this
> is about; ELOOP, EOVERFLOW or EDEADLOCK, maybe?).


The 'allow_redirect' parameter is a very good idea! If I also forward that
to dev_map_bpf_prog_run and do something like the following, I can just
reuse the existing error handling. Or is that too implict/too ugly?

switch (act) {
case XDP_PASS:
        err = xdp_update_frame_from_buff(&xdp, xdpf);
        if (unlikely(err < 0))
                xdp_return_frame_rx_napi(xdpf);
        else
                frames[nframes++] = xdpf;
        break;
case XDP_REDIRECT:
        if (allow_redirect) {
                err = xdp_do_redirect(rx_dev, &xdp, xdp_prog);
                if (unlikely(err))
                        xdp_return_frame_rx_napi(xdpf);
                else
                        *redirects += 1;
                break;
        } else
                fallthrough;
default:
        bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
        fallthrough;
case XDP_ABORTED:
        trace_xdp_exception(tx_dev, xdp_prog, act);
        fallthrough;
case XDP_DROP:
        xdp_return_frame_rx_napi(xdpf);
        break;
}


> 
>>> In any case, this needs extensive selftests, including ones with devmap
>>> programs that loop packets (both redirect from a->a, and from
>>> a->b->a->b) to make sure the limits work correctly.
>>
>>
>> Good point! I am going to prepare some.
>>
>>
>>>
>>>> +	}
>>>>  
>>>>  	/* Ingress dev_rx will be the same for all xdp_frame's in
>>>>  	 * bulk_queue, because bq stored per-CPU and must be flushed
>>>> diff --git a/net/core/filter.c b/net/core/filter.c
>>>> index 8569cd2482ee..b33fc0b1444a 100644
>>>> --- a/net/core/filter.c
>>>> +++ b/net/core/filter.c
>>>> @@ -4287,14 +4287,18 @@ static const struct bpf_func_proto bpf_xdp_adjust_meta_proto = {
>>>>  void xdp_do_flush(void)
>>>>  {
>>>>  	struct list_head *lh_map, *lh_dev, *lh_xsk;
>>>> +	int redirect;
>>>>  
>>>> -	bpf_net_ctx_get_all_used_flush_lists(&lh_map, &lh_dev, &lh_xsk);
>>>> -	if (lh_dev)
>>>> -		__dev_flush(lh_dev);
>>>> -	if (lh_map)
>>>> -		__cpu_map_flush(lh_map);
>>>> -	if (lh_xsk)
>>>> -		__xsk_map_flush(lh_xsk);
>>>> +	do {
>>>> +		redirect = 0;
>>>> +		bpf_net_ctx_get_all_used_flush_lists(&lh_map, &lh_dev, &lh_xsk);
>>>> +		if (lh_dev)
>>>> +			__dev_flush(lh_dev, &redirect);
>>>> +		if (lh_map)
>>>> +			__cpu_map_flush(lh_map);
>>>> +		if (lh_xsk)
>>>> +			__xsk_map_flush(lh_xsk);
>>>> +	} while (redirect > 0);
>>>>  }
>>>>  EXPORT_SYMBOL_GPL(xdp_do_flush);
>>>>  
>>>> @@ -4303,20 +4307,24 @@ void xdp_do_check_flushed(struct napi_struct *napi)
>>>>  {
>>>>  	struct list_head *lh_map, *lh_dev, *lh_xsk;
>>>>  	bool missed = false;
>>>> +	int redirect;
>>>>  
>>>> -	bpf_net_ctx_get_all_used_flush_lists(&lh_map, &lh_dev, &lh_xsk);
>>>> -	if (lh_dev) {
>>>> -		__dev_flush(lh_dev);
>>>> -		missed = true;
>>>> -	}
>>>> -	if (lh_map) {
>>>> -		__cpu_map_flush(lh_map);
>>>> -		missed = true;
>>>> -	}
>>>> -	if (lh_xsk) {
>>>> -		__xsk_map_flush(lh_xsk);
>>>> -		missed = true;
>>>> -	}
>>>> +	do {
>>>> +		redirect = 0;
>>>> +		bpf_net_ctx_get_all_used_flush_lists(&lh_map, &lh_dev, &lh_xsk);
>>>> +		if (lh_dev) {
>>>> +			__dev_flush(lh_dev, &redirect);
>>>> +			missed = true;
>>>> +		}
>>>> +		if (lh_map) {
>>>> +			__cpu_map_flush(lh_map);
>>>> +			missed = true;
>>>> +		}
>>>> +		if (lh_xsk) {
>>>> +			__xsk_map_flush(lh_xsk);
>>>> +			missed = true;
>>>> +		}
>>>> +	} while (redirect > 0);
>>>
>>> With the change suggested above (so that bq_xmit_all() guarantees the
>>> flush is completely done), this looping is not needed anymore. However,
>>> it becomes important in which *order* the flushing is done
>>> (__dev_flush() should always happen first), so adding a comment to note
>>> this would be good.
>>
>>
>> I guess, if we remove the loop here and we still want to cover the case of
>> redirecting from devmap program via cpumap, we need to fetch the lh_map again
>> after calling __dev_flush, right?
>> So I think we should no longer use bpf_net_ctx_get_all_used_flush_lists then:
>>
>>         lh_dev = bpf_net_ctx_get_dev_flush_list();
>>         if (lh_dev)
>>                 __dev_flush(lh_dev);
>>         lh_map = bpf_net_ctx_get_cpu_map_flush_list();
>>         if (lh_map)
>>                 __cpu_map_flush(lh_map);
>>         lh_xsk = bpf_net_ctx_get_xskmap_flush_list();
>>         if (lh_xsk)
>>                 __xsk_map_flush(lh_xsk);
>>
>> But bpf_net_ctx_get_all_used_flush_lists also includes additional checks
>> for IS_ENABLED(CONFIG_BPF_SYSCALL) and IS_ENABLED(CONFIG_XDP_SOCKETS),
>> so I guess they should be directly included in the xdp_do_flush and
>> xdp_do_check_flushed?
>> Then we could remove the bpf_net_ctx_get_all_used_flush_lists.
>> Or do you have an idea for a more elegant solution?
> 
> I think cpumap is fine because that doesn't immediately redirect back to
> the bulk queue; instead __cpu_map_flush() will put the frames on a
> per-CPU ring buffer and wake the kworker on that CPU. Which will in turn
> do another xdp_do_flush() if the cpumap program does a redirect. So in
> other words, we only need to handle recursion of devmap redirects :)

I likely miss something here, but the scenario I am thinking about is the following:

1. cpu_map and xsk_map flush list are empty, while the dev flush map has
   a single frame pending, i.e. in the current implementation after
   executing bpf_net_ctx_get_all_used_flush_lists:
   lh_dev = something
   lh_map = lh_xsk = NULL

2. __dev_flush gets called and executes a bpf program on the frame
   that returns with "return bpf_redirect_map(&cpu_map, 0, 0);"
   and that adds an item to the cpumap flush list.

3. Since __dev_flush is only able to handle devmap redirects itself,
   the item is still on the cpumap flush list after __dev_flush
   has returned.

4. lh_map, however, is still NULL, so __cpu_map_flush does not
   get called and thus the kworker is never woken up.

That is at least what I see on the running system that the kworker
is never woken up, but I might have done something else wrong...

Thanks,
Florian