Content-Type: multipart/mixed; boundary="===============6072256835264964098==" MIME-Version: 1.0 From: Christoph Paasch To: mptcp at lists.01.org Subject: Re: [MPTCP] [PATCH 1/2] skbuff: Add shared control buffer Date: Sun, 29 Oct 2017 21:16:40 -0700 Message-ID: <20171030041640.GI29152@da0602a-dhcp227.apple.com> In-Reply-To: alpine.OSX.2.21.1710231411170.5863@mjmartin-mac01.sea.intel.com X-Status: X-Keywords: X-UID: 143 --===============6072256835264964098== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On 23/10/17 - 15:51:26, Mat Martineau wrote: > = > Hi Christoph, > = > On Mon, 23 Oct 2017, Christoph Paasch wrote: > = > > Hello Mat, > > = > > On 20/10/17 - 16:02:31, Mat Martineau wrote: > > > The sk_buff control buffer is of limited size, and cannot be enlarged > > > without significant impact on systemwide memory use. However, additio= nal > > > per-packet state is needed for some protocols, like Multipath TCP. > > > = > > > An optional shared control buffer placed after the normal struct > > > skb_shared_info can accomodate the necessary state without imposing > > > extra memory usage or code changes on normal struct sk_buff > > > users. __alloc_skb will now place a skb_shared_info_ext structure at > > > skb->end when given the SKB_ALLOC_SHINFO_EXT flag. Most users of stru= ct > > > sk_buff continue to use the skb_shinfo() macro to access shared > > > info. skb_shinfo(skb)->is_ext is set if the extended structure is > > > available, and cleared if it is not. > > > = > > > pskb_expand_head will preserve the shared control buffer if it is pre= sent. > > > = > > > Signed-off-by: Mat Martineau > > > --- > > > include/linux/skbuff.h | 24 +++++++++++++++++++++- > > > net/core/skbuff.c | 56 ++++++++++++++++++++++++++++++++++++++--= ---------- > > > 2 files changed, 66 insertions(+), 14 deletions(-) > > = > > While digging through your patch, I realized that there could be one is= sue. > > = > > In many places in the code we use sizeof(struct skb_shared_info) to com= pute > > the overhead (e.g., see tcp_sndbuf_expand). > > = > > There are also countless users of (sizeof(struct skb_shared_info) in the > > drivers. They all seem to be in the rx-path, so it should be fine. > > = > > But the prevalent use of skb_shared_info throughout the stack seems a b= it > > scary to me. My concern here is that the driver assumes that the overhe= ad is > > always sizeof(skb_shared_info) and thus underestimates the size of the = skb. > > = > > At least, in tcp_sndbuf_expand() the overhead-estimation would be wrong= . I > > don't think the consequences will be catastrophic, but it's something to > > think about. > > = > > = > > What do you think? > > = > = > I saw two typical cases when I looked around existing code: > = > 1. Drivers doing calculations before calling build_skb() on the rx path. > = > 2. Various code making size estimates before creating skbs. > = > = > For #1, build_skb() is still creating "normal" skbs with skb_shared_info,= so > their calculations remain valid. > = > Similarly, code that uses size estimates for #2 goes on to allocate "norm= al" > skbs, so their estimates remains valid. Code that allocates extended skbs > would need another set of extended skb macros to make better estimates. > = > tcp_sndbuf_expand() is a little different, since it changes sk_sndbuf whi= ch > could impact behavior when longer skbs are in use. In this case, it still > seems like a fairly coarse estimate (rounding up to a power of two and th= en > multiplying). I'm not worried about this use, but I do want to understand= it > better before upstreaming. If it's important to add in the extra space for > sockets using skb_shared_info_ext, that could be accounted for in > tcp_sndbuf_expand by adding a flag to tcp_sock. I agree that tcp_sndbuf_expand() should be fine. > Anywhere that allocates extended skbs will involve new code, and if that > code needs to make size estimates it should take the extra overhead in to > account. For existing code that may handle (rather than allocate) skbs wi= th > extended shared data, these skbs will be truthful about their truesize and > will be freed correctly. > = > I still want to review this approach carefully and do additional testing.= I > should run the patch by some network driver authors too. How do you want to proceed for submitting it upstream? I fear that without a user of skb_shared_info_ext, netdev won't accept it. Shouldn't we try to find a user that would benefit from it? Btw., I am trying to get the MD5-changes/extra-option changes ready. But I got hooked up in other pieces last week... Christoph > = > = > Thanks, > Mat > = > = > > > = > > > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h > > > index 03634ec2f918..873910c66df9 100644 > > > --- a/include/linux/skbuff.h > > > +++ b/include/linux/skbuff.h > > > @@ -489,7 +489,8 @@ int skb_zerocopy_iter_stream(struct sock *sk, str= uct sk_buff *skb, > > > * the end of the header data, ie. at skb->end. > > > */ > > > struct skb_shared_info { > > > - __u8 __unused; > > > + __u8 is_ext:1, > > > + __unused:7; > > > __u8 meta_len; > > > __u8 nr_frags; > > > __u8 tx_flags; > > > @@ -530,6 +531,24 @@ struct skb_shared_info { > > > #define SKB_DATAREF_MASK ((1 << SKB_DATAREF_SHIFT) - 1) > > > = > > > = > > > +/* This is an extended version of skb_shared_info, also invariant ac= ross > > > + * clones and living at the end of the header data. > > > + */ > > > +struct skb_shared_info_ext { > > > + /* skb_shared_info must be the first member */ > > > + struct skb_shared_info shinfo; > > > + > > > + /* This is the shared control buffer. It is similar to sk_buff's > > > + * control buffer, but is shared across clones. It must not be > > > + * modified when multiple sk_buffs are referencing this structure. > > > + */ > > > + char shcb[48]; > > > +}; > > > + > > > +#define SKB_SHINFO_EXT_OVERHEAD \ > > > + (SKB_DATA_ALIGN(sizeof(struct skb_shared_info_ext)) - \ > > > + SKB_DATA_ALIGN(sizeof(struct skb_shared_info))) > > > + > > > enum { > > > SKB_FCLONE_UNAVAILABLE, /* skb has no fclone (from head_cache) */ > > > SKB_FCLONE_ORIG, /* orig skb (from fclone_cache) */ > > > @@ -856,6 +875,7 @@ struct sk_buff { > > > #define SKB_ALLOC_FCLONE 0x01 > > > #define SKB_ALLOC_RX 0x02 > > > #define SKB_ALLOC_NAPI 0x04 > > > +#define SKB_ALLOC_SHINFO_EXT 0x08 > > > = > > > /* Returns true if the skb was allocated from PFMEMALLOC reserves */ > > > static inline bool skb_pfmemalloc(const struct sk_buff *skb) > > > @@ -1271,6 +1291,8 @@ static inline unsigned int skb_end_offset(const= struct sk_buff *skb) > > > = > > > /* Internal */ > > > #define skb_shinfo(SKB) ((struct skb_shared_info *)(skb_end_pointer(= SKB))) > > > +#define skb_shinfo_ext(SKB) \ > > > + ((struct skb_shared_info_ext *)(skb_end_pointer(SKB))) > > > = > > > static inline struct skb_shared_hwtstamps *skb_hwtstamps(struct sk_b= uff *skb) > > > { > > > diff --git a/net/core/skbuff.c b/net/core/skbuff.c > > > index 40717501cbdd..397edd5c0613 100644 > > > --- a/net/core/skbuff.c > > > +++ b/net/core/skbuff.c > > > @@ -166,6 +166,8 @@ static void *__kmalloc_reserve(size_t size, gfp_t= flags, int node, > > > * instead of head cache and allocate a cloned (child) skb. > > > * If SKB_ALLOC_RX is set, __GFP_MEMALLOC will be used for > > > * allocations in case the data is required for writeback > > > + * If SKB_ALLOC_SHINFO_EXT is set, the skb will be allocated > > > + * with an extended shared info struct. > > > * @node: numa node to allocate memory on > > > * > > > * Allocate a new &sk_buff. The returned buffer has no headroom and a > > > @@ -179,9 +181,9 @@ struct sk_buff *__alloc_skb(unsigned int size, gf= p_t gfp_mask, > > > int flags, int node) > > > { > > > struct kmem_cache *cache; > > > - struct skb_shared_info *shinfo; > > > struct sk_buff *skb; > > > u8 *data; > > > + unsigned int shinfo_size; > > > bool pfmemalloc; > > > = > > > cache =3D (flags & SKB_ALLOC_FCLONE) > > > @@ -199,18 +201,22 @@ struct sk_buff *__alloc_skb(unsigned int size, = gfp_t gfp_mask, > > > /* We do our best to align skb_shared_info on a separate cache > > > * line. It usually works because kmalloc(X > SMP_CACHE_BYTES) gives > > > * aligned memory blocks, unless SLUB/SLAB debug is enabled. > > > - * Both skb->head and skb_shared_info are cache line aligned. > > > + * Both skb->head and skb_shared_info (or skb_shared_info_ext) are > > > + * cache line aligned. > > > */ > > > size =3D SKB_DATA_ALIGN(size); > > > - size +=3D SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); > > > - data =3D kmalloc_reserve(size, gfp_mask, node, &pfmemalloc); > > > + if (flags & SKB_ALLOC_SHINFO_EXT) > > > + shinfo_size =3D SKB_DATA_ALIGN(sizeof(struct skb_shared_info_ext)); > > > + else > > > + shinfo_size =3D SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); > > > + data =3D kmalloc_reserve(size + shinfo_size, gfp_mask, node, &pfmem= alloc); > > > if (!data) > > > goto nodata; > > > /* kmalloc(size) might give us more room than requested. > > > * Put skb_shared_info exactly at the end of allocated zone, > > > * to allow max possible filling before reallocation. > > > */ > > > - size =3D SKB_WITH_OVERHEAD(ksize(data)); > > > + size =3D ksize(data) - shinfo_size; > > > prefetchw(data + size); > > > = > > > /* > > > @@ -220,7 +226,10 @@ struct sk_buff *__alloc_skb(unsigned int size, g= fp_t gfp_mask, > > > */ > > > memset(skb, 0, offsetof(struct sk_buff, tail)); > > > /* Account for allocated memory : skb + skb->head */ > > > - skb->truesize =3D SKB_TRUESIZE(size); > > > + if (flags & SKB_ALLOC_SHINFO_EXT) > > > + skb->truesize =3D SKB_TRUESIZE(size) + SKB_SHINFO_EXT_OVERHEAD; > > > + else > > > + skb->truesize =3D SKB_TRUESIZE(size); > > > skb->pfmemalloc =3D pfmemalloc; > > > refcount_set(&skb->users, 1); > > > skb->head =3D data; > > > @@ -231,10 +240,21 @@ struct sk_buff *__alloc_skb(unsigned int size, = gfp_t gfp_mask, > > > skb->transport_header =3D (typeof(skb->transport_header))~0U; > > > = > > > /* make sure we initialize shinfo sequentially */ > > > - shinfo =3D skb_shinfo(skb); > > > - memset(shinfo, 0, offsetof(struct skb_shared_info, dataref)); > > > - atomic_set(&shinfo->dataref, 1); > > > - kmemcheck_annotate_variable(shinfo->destructor_arg); > > > + if (flags & SKB_ALLOC_SHINFO_EXT) { > > > + struct skb_shared_info_ext *shinfo_ext =3D skb_shinfo_ext(skb); > > > + shinfo_ext->shinfo.is_ext =3D 1; > > > + memset(&shinfo_ext->shinfo.meta_len, 0, > > > + offsetof(struct skb_shared_info, dataref) - > > > + offsetof(struct skb_shared_info, meta_len)); > > > + atomic_set(&shinfo_ext->shinfo.dataref, 1); > > > + kmemcheck_annotate_variable(shinfo_ext->shinfo.destructor_arg); > > > + memset(&shinfo_ext->shcb, 0, sizeof(shinfo_ext->shcb)); > > > + } else { > > > + struct skb_shared_info *shinfo =3D skb_shinfo(skb); > > > + memset(shinfo, 0, offsetof(struct skb_shared_info, dataref)); > > > + atomic_set(&shinfo->dataref, 1); > > > + kmemcheck_annotate_variable(shinfo->destructor_arg); > > > + } > > > = > > > if (flags & SKB_ALLOC_FCLONE) { > > > struct sk_buff_fclones *fclones; > > > @@ -1443,6 +1463,7 @@ int pskb_expand_head(struct sk_buff *skb, int n= head, int ntail, > > > { > > > int i, osize =3D skb_end_offset(skb); > > > int size =3D osize + nhead + ntail; > > > + int shinfo_size; > > > long off; > > > u8 *data; > > > = > > > @@ -1454,11 +1475,14 @@ int pskb_expand_head(struct sk_buff *skb, int= nhead, int ntail, > > > = > > > if (skb_pfmemalloc(skb)) > > > gfp_mask |=3D __GFP_MEMALLOC; > > > - data =3D kmalloc_reserve(size + SKB_DATA_ALIGN(sizeof(struct skb_sh= ared_info)), > > > - gfp_mask, NUMA_NO_NODE, NULL); > > > + if (skb_shinfo(skb)->is_ext) > > > + shinfo_size =3D SKB_DATA_ALIGN(sizeof(struct skb_shared_info_ext)); > > > + else > > > + shinfo_size =3D SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); > > > + data =3D kmalloc_reserve(size + shinfo_size, gfp_mask, NUMA_NO_NODE= , NULL); > > > if (!data) > > > goto nodata; > > > - size =3D SKB_WITH_OVERHEAD(ksize(data)); > > > + size =3D ksize(data) - shinfo_size; > > > = > > > /* Copy only real data... and, alas, header. This should be > > > * optimized for the cases when header is void. > > > @@ -1468,6 +1492,12 @@ int pskb_expand_head(struct sk_buff *skb, int = nhead, int ntail, > > > memcpy((struct skb_shared_info *)(data + size), > > > skb_shinfo(skb), > > > offsetof(struct skb_shared_info, frags[skb_shinfo(skb)->nr_f= rags])); > > > + if (skb_shinfo(skb)->is_ext) { > > > + int offset =3D offsetof(struct skb_shared_info_ext, shcb); > > > + memcpy((struct skb_shared_info_ext *)(data + size + offset), > > > + &skb_shinfo_ext(skb)->shcb, > > > + sizeof(skb_shinfo_ext(skb)->shcb)); > > > + } > > > = > > > /* > > > * if shinfo is shared we must drop the old head gracefully, but if= it > > > -- > > > 2.14.2 > > > = > > > _______________________________________________ > > > mptcp mailing list > > > mptcp(a)lists.01.org > > > https://lists.01.org/mailman/listinfo/mptcp > > = > = > -- > Mat Martineau > Intel OTC --===============6072256835264964098==--