Re: [Xen-devel] Xen-unstable Linux 3.14-rc3 and 3.13 Network troubles "bisected"

[Sander Eikelenboom <linux@xxxxxxxxxxxxxx>]

Wednesday, March 26, 2014, 3:44:42 PM, you wrote:

>> -----Original Message-----
>> From: Sander Eikelenboom [mailto:linux@xxxxxxxxxxxxxx]
>> Sent: 26 March 2014 11:11
>> To: Paul Durrant
>> Cc: Wei Liu; annie li; Zoltan Kiss; xen-devel@xxxxxxxxxxxxx; Ian Campbell; 
>> linux-
>> kernel; netdev@xxxxxxxxxxxxxxx
>> Subject: Re: [Xen-devel] Xen-unstable Linux 3.14-rc3 and 3.13 Network
>> troubles "bisected"
>> 
>> Paul,
>> 
>> You have been awfully silent for this whole thread while this is a regression
>> caused by a patch of you
>> (ca2f09f2b2c6c25047cfc545d057c4edfcfe561c as clearly stated much earlier in
>> this thread).
>> 

> Sorry, I've been distracted...

>> The commit messages states:
>>     "net_rx_action() is the place where we could do with an accurate
>> predicition but,
>>     since that has proven tricky to calculate, a cheap worse-case (but not 
>> too
>> bad)
>>     estimate is all we really need since the only thing we *must* prevent is
>> xenvif_gop_skb()
>>     consuming more slots than are available."
>> 
>> Your "worst-case" calculation stated in the commit message is clearly not the
>> worst case,
>> since it doesn't take calls to "get_next_rx_buffer" into account.
>> 

> It should be taking into account the behaviour of start_new_rx_buffer(), 
> which should be true if a slot is full or a frag will overflow the current 
> slot and doesn't require splitting.
> The code in net_rx_action() makes the assumption that each frag will require 
> as many slots as its size requires, i.e. it assumes no packing of multiple 
> frags into a single slot, so it should be a worst case.
> Did I miss something in that logic?

Yes.
In "xenvif_gop_skb()" this loop:

        for (i = 0; i < nr_frags; i++) {
                xenvif_gop_frag_copy(vif, skb, npo,
                                     skb_frag_page(&skb_shinfo(skb)->frags[i]),
                                     skb_frag_size(&skb_shinfo(skb)->frags[i]),
                                     skb_shinfo(skb)->frags[i].page_offset,
                                     &head);
        }

Is capable of using up (at least) 1 slot more than is anticipated for in 
"net_rx_action()"  by this code:

                for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
                        unsigned int size;
                        size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
                        max_slots_needed += DIV_ROUND_UP(size, PAGE_SIZE);
                }

And this happens when it calls "get_next_rx_buffer()" from 
"xenvif_gop_frag_copy()" where it's breaking down the frag.

Ultimately this results in bad grant reference warnings (and packets marked as 
"errors" in the interface statistics).

In my case it always seems to be a skb with 1 frag which is broken down in 5 or 
6 pieces ..

So "get_next_rx_buffer()" is called once .. and i'm overrunning the ring with 1 
slot, but i'm not sure if that's not coincedence
since in the code there seem to be no explicit limitation on how often this 
code path is taken. So perhaps it's implicitly limited
since packets and frags can't be arbitrarily large in comparison with the 
page_size but that's not something i'm capable of figuring out :-)



>   Paul

>> Problem is that a worst case calculation would probably be reverting to the
>> old calculation,
>> and the problems this patch was trying to solve would reappear, but
>> introducing new regressions
>> isn't very useful either. And since it seems such a tricky and fragile thing 
>> to
>> determine, it would
>> probably be best to be split into a distinct function with a comment to 
>> explain
>> the rationale used.
>> 
>> Since this doesn't seem to progress very fast .. CC'ed some more folks .. you
>> never know ..
>> 
>> --
>> Sander
>> 
>> 
>> Tuesday, March 25, 2014, 4:29:42 PM, you wrote:
>> 
>> 
>> > Tuesday, March 25, 2014, 4:15:39 PM, you wrote:
>> 
>> >> On Sat, Mar 22, 2014 at 07:28:34PM +0100, Sander Eikelenboom wrote:
>> >> [...]
>> >>> > Yes there is only one frag .. but it seems to be much larger than
>> PAGE_SIZE .. and xenvif_gop_frag_copy brakes that frag down into smaller
>> bits .. hence the calculation in xenvif_rx_action determining the slots 
>> needed
>> by doing:
>> >>>
>> >>> >                 for (i = 0; i < nr_frags; i++) {
>> >>> >                         unsigned int size;
>> >>> >                         size = 
>> >>> > skb_frag_size(&skb_shinfo(skb)->frags[i]);
>> >>> >                         max_slots_needed += DIV_ROUND_UP(size, 
>> >>> > PAGE_SIZE);
>> >>> >                 }
>> >>>
>> >>> > But the code in xenvif_gop_frag_copy .. seems to be needing one
>> more slot (from the emperical test) .. and calling "get_next_rx_buffer"
>> seems involved in that ..
>> >>>
>> >>> Hmm looked again .. and it seems this is it .. when your frags are large
>> enough you have the chance of running into this.
>> >>>
>> 
>> >> get_next_rx_buffer is guarded by start_new_rx_buffer. Do you see any
>> >> problem with that implementation?
>> > In general no, but "get_next_rx_buffer" up's cons .. and the calculations
>> done in "xenvif_rx_action" for max_slots_needed to prevent the overrun
>> > don't count in this possibility. So it's not the guarding of
>> "start_new_rx_buffer" that is at fault. It's the ones early in
>> "xenvif_rx_action".
>> > The ones that were changed by Paul's patch from MAX_SKB_FRAGS to a
>> calculated value that should be a "slim fit".
>> 
>> > The problem is in determining upfront in "xenvif_rx_action" when and how
>> often the "get_next_rx_buffer" path will be taken.
>> > Unless there are other non direct restrictions (from a size point of view) 
>> > it
>> can be called multiple times per frag per skb.
>> 
>> >>> Problem is .. i don't see an easy fix, the "one more slot" of the 
>> >>> empirical
>> test doesn't seem to be the worst case either (i think):
>> >>>
>> >>> - In my case the packets that hit this only have 1 frag, but i could have
>> had more frags.
>> >>>   I also think you can't rule out the possibility of doing the
>> "get_next_rx_buffer" for multiple subsequent frags from one packet,
>> >>>   so in the worst (and perhaps even from a single frag since it's looped
>> over a split of it in what seems PAGE_SIZE pieces.)
>> >>>   - So an exact calculation of how much slots we are going to need for
>> hitting this "get_next_rx_buffer"  upfront in "xenvif_rx_action" seems
>> unfeasible.
>> >>>   - A worst case gamble seems impossible either .. if you take multiple
>> frags * multiple times the "get_next_rx_buffer" ... you would probably be
>> back at just
>> >>>     setting the needed_slots to MAX_SKB_FRAGS.
>> >>>
>> >>> - Other thing would be checking for the available slots before doing the
>> "get_next_rx_buffer" .. how ever .. we don't account for how many slots we
>> still need to
>> >>>   just process the remaining frags.
>> >>>
>> 
>> >> We've done a worst case estimation for whole SKB (linear area + all
>> >> frags) already, at the very first beginning. That's what
>> >> max_slots_needed is for.
>> 
>> >>> - Just remove the whole "get_next_rx_buffer".. just tested it but it
>> seems the "get_next_rx_buffer" is necessary ..  when i make
>> start_new_rx_buffer always return false
>> >>>   i hit the bug below :S
>> >>>
>> >>> So the questions are ... is the "get_next_rx_buffer" part really 
>> >>> necessary
>> ?
>> >>>    - if not, what is the benefit of the "get_next_rx_buffer" and does 
>> >>> that
>> outweigh the additional cost of accounting
>> >>>      of needed_slots for the frags that have yet to be processed ?
>> >>>    - if yes, erhmmm what would be the best acceptable solution ..
>> returning to using MAX_SKB_FRAGS ?
>> >>>
>> 
>> >> I think you need to answer several questions first:
>> >> 1. is max_slots_needed actually large enough to cover whole SKB?
>> >         No it's not if, you end up calling "get_next_rx_buffer" one or 
>> > multiple
>> times when processing the SKB
>> >         you have the chance of overrunning (or be saved because prod gets
>> upped before you overrun).
>> 
>> >> 2. is the test of ring slot availability acurate?
>> >         Seems to be.
>> 
>> >> 3. is the number of ring slots consumed larger than max_slots_needed? (I
>> >>    guess the answer is yes)
>> >         Yes that was the whole point.
>> 
>> >> 4. which step in the break-down procedure causes backend to overrun
>> the
>> >>    ring?
>> >         The "get_next_rx_buffer" call .. that is not accounted for (which 
>> > can be
>> called
>> >         multiple times per frag (unless some other none obvious size
>> restriction limits this
>> >         to one time per frag or one time per SKB).
>> >         In my errorneous case it is called one time, but it would be nice 
>> > if there
>> would be some theoretical proof
>> >         instead of just some emperical test.
>> 
>> 
>> >> It doesn't matter how many frags your SKB has and how big a frag is. If
>> >> every step is correct then you're fine. The code you're looking at
>> >> (start_new_rx_buffer / get_next_rx_buffer and friend) needs to be
>> >> carefully examined.
>> 
>> > Well it seems it only calls "get_next_rx_buffer" in some special cases ..
>> (and that also what i'm seeing because it doesn't happen
>> > continously.
>> 
>> > Question is shouldn't this max_needed_slots calc be reverted to what it
>> was before 3.14 and take the time in 3.15 to figure out a
>> > the theoretical max (if it can be calculated upfront) .. or another way to
>> guarantee the code is able to process the whole SKB  ?
>> 
>> >> Wei.
>> 
>> 
>> 
>> 




_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel