[dpdk-dev] [PATCH v2 1/6] mbuf: add buffer offset field for flexible indirection

[Yongseok Koh]

On Wed, Apr 11, 2018 at 11:39:47AM +0000, Ananyev, Konstantin wrote:
> 
> Hi Yongseok,
> 
> > > >
> > > > On Mon, Apr 09, 2018 at 06:04:34PM +0200, Olivier Matz wrote:
> > > > > Hi Yongseok,
> > > > >
> > > > > On Tue, Apr 03, 2018 at 05:12:06PM -0700, Yongseok Koh wrote:
> > > > > > On Tue, Apr 03, 2018 at 10:26:15AM +0200, Olivier Matz wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > On Mon, Apr 02, 2018 at 11:50:03AM -0700, Yongseok Koh wrote:
> > > > > > > > When attaching a mbuf, indirect mbuf has to point to start of buffer of
> > > > > > > > direct mbuf. By adding buf_off field to rte_mbuf, this becomes more
> > > > > > > > flexible. Indirect mbuf can point to any part of direct mbuf by calling
> > > > > > > > rte_pktmbuf_attach_at().
> > > > > > > >
> > > > > > > > Possible use-cases could be:
> > > > > > > > - If a packet has multiple layers of encapsulation, multiple indirect
> > > > > > > >   buffers can reference different layers of the encapsulated packet.
> > > > > > > > - A large direct mbuf can even contain multiple packets in series and
> > > > > > > >   each packet can be referenced by multiple mbuf indirections.
> > > > > > > >
> > > > > > > > Signed-off-by: Yongseok Koh <yskoh at mellanox.com>
> > > > > > >
> > > > > > > I think the current API is already able to do what you want.
> > > > > > >
> > > > > > > 1/ Here is a mbuf m with its data
> > > > > > >
> > > > > > >                off
> > > > > > >                <-->
> > > > > > >                       len
> > > > > > >           +----+   <---------->
> > > > > > >           |    |
> > > > > > >         +-|----v----------------------+
> > > > > > >         | |    -----------------------|
> > > > > > > m       | buf  |    XXXXXXXXXXX      ||
> > > > > > >         |      -----------------------|
> > > > > > >         +-----------------------------+
> > > > > > >
> > > > > > >
> > > > > > > 2/ clone m:
> > > > > > >
> > > > > > >   c = rte_pktmbuf_alloc(pool);
> > > > > > >   rte_pktmbuf_attach(c, m);
> > > > > > >
> > > > > > >   Note that c has its own offset and length fields.
> > > > > > >
> > > > > > >
> > > > > > >                off
> > > > > > >                <-->
> > > > > > >                       len
> > > > > > >           +----+   <---------->
> > > > > > >           |    |
> > > > > > >         +-|----v----------------------+
> > > > > > >         | |    -----------------------|
> > > > > > > m       | buf  |    XXXXXXXXXXX      ||
> > > > > > >         |      -----------------------|
> > > > > > >         +------^----------------------+
> > > > > > >                |
> > > > > > >           +----+
> > > > > > > indirect  |
> > > > > > >         +-|---------------------------+
> > > > > > >         | |    -----------------------|
> > > > > > > c       | buf  |                     ||
> > > > > > >         |      -----------------------|
> > > > > > >         +-----------------------------+
> > > > > > >
> > > > > > >                 off    len
> > > > > > >                 <--><---------->
> > > > > > >
> > > > > > >
> > > > > > > 3/ remove some data from c without changing m
> > > > > > >
> > > > > > >    rte_pktmbuf_adj(c, 10)   // at head
> > > > > > >    rte_pktmbuf_trim(c, 10)  // at tail
> > > > > > >
> > > > > > >
> > > > > > > Please let me know if it fits your needs.
> > > > > >
> > > > > > No, it doesn't.
> > > > > >
> > > > > > Trimming head and tail with the current APIs removes data and make the space
> > > > > > available. Adjusting packet head means giving more headroom, not shifting the
> > > > > > buffer itself. If m has two indirect mbufs (c1 and c2) and those are pointing to
> > > > > > difference offsets in m,
> > > > > >
> > > > > > rte_pktmbuf_adj(c1, 10);
> > > > > > rte_pktmbuf_adj(c2, 20);
> > > > > >
> > > > > > then the owner of c2 regard the first (off+20)B as available headroom. If it
> > > > > > wants to attach outer header, it will overwrite the headroom even though the
> > > > > > owner of c1 is still accessing it. Instead, another mbuf (h1) for the outer
> > > > > > header should be linked by h1->next = c2.
> > > > >
> > > > > Yes, after these operations c1, c2 and m should become read-only. So, to
> > > > > prepend headers, another mbuf has to be inserted before as you suggest. It
> > > > > is possible to wrap this in a function rte_pktmbuf_clone_area(m, offset,
> > > > > length) that will:
> > > > >   - alloc and attach indirect mbuf for each segment of m that is
> > > > >     in the range [offset : length+offset].
> > > > >   - prepend an empty and writable mbuf for the headers
> > > > >
> > > > > > If c1 and c2 are attached with shifting buffer address by adjusting buf_off,
> > > > > > which actually shrink the headroom, this case can be properly handled.
> > > > >
> > > > > What do you mean by properly handled?
> > > > >
> > > > > Yes, prepending data or adding data in the indirect mbuf won't override
> > > > > the direct mbuf. But prepending data or adding data in the direct mbuf m
> > > > > won't be protected.
> > > > >
> > > > > From an application point of view, indirect mbufs, or direct mbufs that
> > > > > have refcnt != 1, should be both considered as read-only because they
> > > > > may share their data. How an application can know if the data is shared
> > > > > or not?
> > > > >
> > > > > Maybe we need a flag to differentiate mbufs that are read-only
> > > > > (something like SHARED_DATA, or simply READONLY). In your case, if my
> > > > > understanding is correct, you want to have indirect mbufs with RW data.
> > > >
> > > > Agree that indirect mbuf must be treated as read-only, Then the current code is
> > > > enough to handle that use-case.
> > > >
> > > > > > And another use-case (this is my actual use-case) is to make a large mbuf have
> > > > > > multiple packets in series. AFAIK, this will also be helpful for some FPGA NICs
> > > > > > because it transfers multiple packets to a single large buffer to reduce PCIe
> > > > > > overhead for small packet traffic like the Multi-Packet Rx of mlx5 does.
> > > > > > Otherwise, packets should be memcpy'd to regular mbufs one by one instead of
> > > > > > indirect referencing.
> > >
> > > But just to make HW to RX multiple packets into one mbuf,
> > > data_off inside indirect mbuf should be enough, correct?
> > Right. Current max buffer len of mbuf is 64kB (16bits) but it is enough for mlx5
> > to reach to 100Gbps with 64B traffic (149Mpps). I made mlx5 HW put 16 packets in
> > a buffer. So, it needs ~32kB buffer. Having more bits in length fields would be
> > better but 16-bit is good enough to overcome the PCIe Gen3 bottleneck in order
> > to saturate the network link.
> 
> There were few complains that 64KB max is a limitation for some use-cases.
> I am not against increasing it, but I don't think we have free space on first cache-line for that
> without another big rework of mbuf layout. 
> Considering that we need to increase size for buf_len, data_off, data_len, and probably priv_size too. 
> 
> > 
> > > As I understand, what you'd like to achieve with this new field -
> > > ability to manipulate packet boundaries after RX, probably at upper layer.
> > > As Olivier pointed above, that doesn't sound as safe approach - as you have multiple
> > > indirect mbufs trying to modify same direct buffer.
> > 
> > I agree that there's an implication that indirect mbuf or mbuf having refcnt > 1
> > is read-only. What that means, all the entities which own such mbufs have to be
> > aware of that and keep the principle as DPDK can't enforce the rule and there
> > can't be such sanity check. In this sense, HW doesn't violate it because the
> > direct mbuf is injected to HW before indirection. When packets are written by
> > HW, PMD attaches indirect mbufs to the direct mbuf and deliver those to
> > application layer with freeing the original direct mbuf (decrement refcnt by 1).
> > So, HW doesn't touch the direct buffer once it reaches to upper layer.
> 
> Yes, I understand that. But as I can see you introduced functions to adjust head and tail,
> which implies that it should be possible by some entity (upper layer?) to manipulate these
> indirect mbufs.
> And we don't know how exactly it will be done.

That's a valid concern. I can make it private by merging into the _attach_to()
func, or I just can add a comment in the API doc. However, if users are aware
that a mbuf is read-only and we expect them to keep it intact by their own
judgement, they would/should not use those APIs. We can't stop them modifying
content or the buffer itself anyway. Will add more comments of this discussion
regarding read-only mode.

> > The direct buffer will be freed and get available for reuse when all the attached
> > indirect mbufs are freed.
> > 
> > > Though if you really need to do that, why it can be achieved by updating buf_len and priv_size
> > > Fields for indirect mbufs, straight after attach()?
> > 
> > Good point.
> > Actually that was my draft (Mellanox internal) version of this patch :-) But I
> > had to consider a case where priv_size is really given by user. Even though it
> > is less likely, but if original priv_size is quite big, it can't cover entire
> > buf_len. For this, I had to increase priv_size to 32-bit but adding another
> > 16bit field (buf_off) looked more plausible.
> 
> As I remember, we can't have mbufs bigger then 64K,
> so priv_size + buf_len should be always less than 64K, correct?

Can you let me know where I can find the constraint? I checked
rte_pktmbuf_pool_create() and rte_pktmbuf_init() again to not make any mistake
but there's no such limitation.

	elt_size = sizeof(struct rte_mbuf) + (unsigned)priv_size +
		(unsigned)data_room_size;

The max of data_room_size is 64kB, so is priv_size. m->buf_addr starts from 'm +
sizeof(*m) + priv_size' and m->buf_len can't be larger than UINT16_MAX. So,
priv_size couldn't be used for this purpose.

Yongseok

> > > > > >
> > > > > > Does this make sense?
> > > > >
> > > > > I understand the need.
> > > > >
> > > > > Another option would be to make the mbuf->buffer point to an external
> > > > > buffer (not inside the direct mbuf). This would require to add a
> > > > > mbuf->free_cb. See "Mbuf with external data buffer" (page 19) in [1] for
> > > > > a quick overview.
> > > > >
> > > > > [1]
> > > >
> > https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdpdksummit.com%2FArchive%2Fpdf%2F2016Userspace%2FDay01
> > > > -Session05-OlivierMatz-
> > > >
> > Userspace2016.pdf&data=02%7C01%7Cyskoh%40mellanox.com%7Ca5405edb36e445e6540808d59e339a38%7Ca652971c7d2e4d9ba6a4d
> > > > 149256f461b%7C0%7C0%7C636588866861082855&sdata=llw%2BwiY5cC56naOUhBbIg8TKtfFN6VZcIRY5PV7VqZs%3D&reserved=0
> > > > >
> > > > > The advantage is that it does not require the large data to be inside a
> > > > > mbuf (requiring a mbuf structure before the buffer, and requiring to be
> > > > > allocated from a mempool). On the other hand, it is maybe more complex
> > > > > to implement compared to your solution.
> > > >
> > > > I knew that you presented the slides and frankly, I had considered that option
> > > > at first. But even with that option, metadata to store refcnt should also be
> > > > allocated and managed anyway. Kernel also maintains the skb_shared_info at the
> > > > end of the data segment. Even though it could have smaller metadata structure,
> > > > I just wanted to make full use of the existing framework because it is less
> > > > complex as you mentioned. Given that you presented the idea of external data
> > > > buffer in 2016 and there hasn't been many follow-up discussions/activities so
> > > > far, I thought the demand isn't so big yet thus I wanted to make this patch
> > > > simpler.  I personally think that we can take the idea of external data seg when
> > > > more demands come from users in the future as it would be a huge change and may
> > > > break current ABI/API. When the day comes, I'll gladly participate in the
> > > > discussions and write codes for it if I can be helpful.
> > > >
> > > > Do you think this patch is okay for now?
> > > >
> > > >
> > > > Thanks for your comments,
> > > > Yongseok