Re: Re: Re: [Xen-devel] SetPageForeign in netback

["Grzegorz Miłoś" <gm281@xxxxxxxxx>]

> what about another question:
>
> In netback netif_be_start_xmit()
>        if (!netif->copying_receiver ||
>            ((skb_headlen(skb) + offset_in_page(skb->data)) >= PAGE_SIZE)) {
>                struct sk_buff *nskb = netbk_copy_skb(skb);
> which calls skb_copy_bits()
>
> it means when rx-flip, the skb data and frags are copied to another buffer.
> then the buffer flip with domU's pages.
> so before page flip with domU, copying happens in dom0. Does rx-flip make any 
> sense?

The reason why you have to copy is to make sure that you pull the skb
into a fresh page (which doesn't have any other pieces of data in it).
It's only then that you can flip it to DomU.
It used to be the case that the backend was more efficient in the
rx-flip mode. Some/most skbs copies were avoided by keeping a close
eye on the skb allocation. Since rx-flip is now only/mostly kept for
backward compatibility (rx-copy is the default) the code was
simplified.
Note: with grant table copy you can ask the hypervisor to copy
sub-page regions (in our case you can match the size of the skb)

Cheers
Gr(z)egor(z)
>
>
>
> ======= 2008-06-23 19:22:37 您在来信中写道：=======
>
>>> thanks. Maybe I understand.
>>> Another question
>>> In netback:
>>> static inline void maybe_schedule_tx_action(void)
>>> {
>>>        smp_mb();
>>>        if ((NR_PENDING_REQS < (MAX_PENDING_REQS/2)) &&
>>>            !list_empty(&net_schedule_list))
>>>                tasklet_schedule(&net_tx_tasklet);
>>> }
>>> ... ...
>>>
>>> while
>>> #define NR_PENDING_REQS (MAX_PENDING_REQS - pending_prod + pending_cons)
>>> that is NR_PENDING_REQS = MAX_PENDING_REQS - (pending_prod - pending_cons)
>>>
>>> My question is when NR_PENDING_REQS cannot satisfy the above condition  
>>> "(NR_PENDING_REQS < (MAX_PENDING_REQS/2)"?
>>> why not handle the condition of "NR_PENDING_REQS >= (MAX_PENDING_REQS/2)"?
>>
>>If there are many pending requests (more then MAX/2), new requests
>>will have to wait on frontend-backend rings (a form of congestion
>>control). When some of the pending requests get dealt with (and
>>NR_PENDING_REQS falls below MAX/2) new ones will be accepted.
>>
>>Also, note that net_tx_tasklet doesn't only get scheduled by
>>maybe_schedule_tx_action. There also is:
>>- netbk_tx_pending_timeout: triggered by a timeout timer
>>- netif_idx_release: triggered by a net data page being released
>>The three net_tx_tasklet schedulers guarantee continuous flow of skbs
>>out of netback.
>>
>>Cheers
>>Gr(z)egor(z)
>>
>>
>>>
>>>
>>> ======= 2008-06-19 15:22:58 您在来信中写道：=======
>>>
>>>>The aim is to achieve batching and scheduling of work (to some extent) on
>>>>both the transmit and receive paths. So no vif gets starved, and larger
>>>>batches of packets are handled more efficiently.
>>>>
>>>> -- Keir
>>>>
>>>>
>>>>On 19/6/08 01:40, "Zang Hongyong" <zanghongyong@xxxxxxxxxx> wrote:
>>>>
>>>>> Thanks again!
>>>>> Another question about netback.
>>>>> tasklet is used in both tx and rx. Lets take a look at rx, Before tasklet,
>>>>> packets of all vnifs must be queued together, and in tasklet packet will 
>>>>> be
>>>>> dequeued, and handled to its proper netfront of domU.
>>>>> 1)why not handle packet directly without the overall queue and tasklet 
>>>>> stuff?
>>>>> 2)Is the overall queue and tasklet stuff fair to all vnifs? For example， 
>>>>> when
>>>>> vif1.0 rx, the netback driver put its packet to overall queue and do 
>>>>> tasklet,
>>>>> while in tasklet, packets belonging to other vif maybe handled.
>>>>>   I've noticed when tx, netfront handle packet directly to its proper
>>>>> Ring,Request stuff.
>>>>>
>>>>>
>>>>>
>>>>> ======= 2008-06-19 00:37:40 您在来信中写道：=======
>>>>>
>>>>>>> Many thanks!
>>>>>>> And that is, when tx, after the data page is sent by native Nic driver 
>>>>>>> in
>>>>>>> dom0, the data page will be freed, then netif_page_release() called 
>>>>>>> which
>>>>>>> indicates netback to unmap the page offered by domU, and moves on its tx
>>>>>>> response.
>>>>>>>
>>>>>>> Is that so?
>>>>>>
>>>>>> Correct.
>>>>>>
>>>>>>> If so, how about a bad NIC driver which doen't call free_page() after
>>>>>>> sending data out of machine ?
>>>>>>
>>>>>> Well, it could happen if there was a memory leak in the driver. This
>>>>>> would also be present in non-xenified linux. We are hoping for
>>>>>> bug-free device drivers.
>>>>>>
>>>>>>
>>>>>>> and Why mmap_pages is allocated by
>>>>>>> alloc_empty_pages_and_pagevec(MAX_PENDING_REQS)?
>>>>>>> can  mmap_pages be allocated by alloc_vm_area() and vmalloc_to_page() ??
>>>>>>
>>>>>> alloc_empty_pages_and_pagevec() balloons machine memory frames away
>>>>>> from Dom0, you are therefore left with pseudo-physical page that's not
>>>>>> backed by real memory. You want that, because you'll substitute DomU's
>>>>>> memory frame in it's place. I don't think alloc_vm_area does that. It
>>>>>> would only allocate virtually continuous range of memory.
>>>>>>
>>>>>> Cheers
>>>>>> Gr(z)egor(z)
>>>>>>
>>>>>>>
>>>>>>> Forgive my silly questions above please.
>>>>>>>
>>>>>>>
>>>>>>> ======= 2008-06-18 18:52:27 您在来信中写道：=======
>>>>>>>
>>>>>>>>> hi,
>>>>>>>>>  in netback init mmap_pages,
>>>>>>>>>                SetPageForeign(page, netif_page_release);
>>>>>>>>>  that is, page->index = netif_page_release
>>>>>>>>>  while netif_page_release is a function.
>>>>>>>>
>>>>>>>> netif_page_release is a function, and therefore:
>>>>>>>> page->index = netif_page_release
>>>>>>>> will store netif_page_release function pointer in 'index'
>>>>>>>>
>>>>>>>>>  so what's the meaning of SetPageForeign?
>>>>>>>>
>>>>>>>> Setting a page foreign means that the page is owned by another domain,
>>>>>>>> and that some care needs to be taken when freeing it.
>>>>>>>>
>>>>>>>>>  And when the function  netif_page_release() will be called?
>>>>>>>>
>>>>>>>> Whenever PageForeignDestructor is called (as it calls the destructor
>>>>>>>> function stored in the 'index' field).
>>>>>>>> PageForeignDestructor is called from:
>>>>>>>> __free_pages_ok
>>>>>>>> and
>>>>>>>> free_hot_cold_page
>>>>>>>>
>>>>>>>> Hope this helps.
>>>>>>>>
>>>>>>>> Cheers
>>>>>>>> Gr(z)egor(z)
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> Xen-devel mailing list
>>>>>>>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>>>>>>>> http://lists.xensource.com/xen-devel
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>> = = = = = = = = = = = = = = = = = = = =
>>>>>>> Zang Hongyong
>>>>>>> zanghongyong@xxxxxxxxxx
>>>>>>> 2008-06-18
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> Xen-devel mailing list
>>>>>>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>>>>>>> http://lists.xensource.com/xen-devel
>>>>>>>
>>>>>>>
>>>>>> _______________________________________________
>>>>>> Xen-devel mailing list
>>>>>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>>>>>> http://lists.xensource.com/xen-devel
>>>>>>
>>>>>
>>>>> = = = = = = = = = = = = = = = = = = = =
>>>>> Zang Hongyong
>>>>> zanghongyong@xxxxxxxxxx
>>>>> 2008-06-19
>>>>>
>>>>> _______________________________________________
>>>>> Xen-devel mailing list
>>>>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>>>>> http://lists.xensource.com/xen-devel
>>>>
>>>>
>>>>
>>>>_______________________________________________
>>>>Xen-devel mailing list
>>>>Xen-devel@xxxxxxxxxxxxxxxxxxx
>>>>http://lists.xensource.com/xen-devel
>>>
>>> = = = = = = = = = = = = = = = = = = = =
>>> Zang Hongyong
>>> zanghongyong@xxxxxxxxxx
>>> 2008-06-22
>>>
>>>
>>> _______________________________________________
>>> Xen-devel mailing list
>>> Xen-devel@xxxxxxxxxxxxxxxxxxx
>>> http://lists.xensource.com/xen-devel
>>>
>>>
>>_______________________________________________
>>Xen-devel mailing list
>>Xen-devel@xxxxxxxxxxxxxxxxxxx
>>http://lists.xensource.com/xen-devel
>>
>
> = = = = = = = = = = = = = = = = = = = =
> Zang Hongyong
> zanghongyong@xxxxxxxxxx
> 2008-06-23
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxx
> http://lists.xensource.com/xen-devel
>
>

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel