[XFRM]: Optimize MTU calculation Replace the probing based MTU estimation, which usually takes 2-3 iterations to find a fitting value and may underestimate the MTU, by an exact calculation. Also fix underestimation of the XFRM trailer_len, which causes unnecessary reallocations. Signed-off-by: Patrick McHardy --- commit 02ce45c1de8a2de2f02cc6f3d88583706575142b tree e97b27090af4a3767df52c16ea23ea9851a97c58 parent c543abe080ba68e9040ca819bab2a6ff2ad90c15 author Patrick McHardy Thu, 05 Apr 2007 14:02:07 +0200 committer Patrick McHardy Thu, 05 Apr 2007 14:02:07 +0200 include/net/xfrm.h | 2 +- net/ipv4/esp4.c | 30 ++++++++++++++++++------------ net/ipv6/esp6.c | 22 ++++++++++++---------- net/xfrm/xfrm_state.c | 36 ++++++++---------------------------- 4 files changed, 39 insertions(+), 51 deletions(-) diff --git a/include/net/xfrm.h b/include/net/xfrm.h index 5a00aa8..e144a25 100644 --- a/include/net/xfrm.h +++ b/include/net/xfrm.h @@ -279,7 +279,7 @@ struct xfrm_type xfrm_address_t *(*local_addr)(struct xfrm_state *, xfrm_address_t *); xfrm_address_t *(*remote_addr)(struct xfrm_state *, xfrm_address_t *); /* Estimate maximal size of result of transformation of a dgram */ - u32 (*get_max_size)(struct xfrm_state *, int size); + u32 (*get_mtu)(struct xfrm_state *, int size); }; extern int xfrm_register_type(struct xfrm_type *type, unsigned short family); diff --git a/net/ipv4/esp4.c b/net/ipv4/esp4.c index 5e56139..24238fb 100644 --- a/net/ipv4/esp4.c +++ b/net/ipv4/esp4.c @@ -272,32 +272,34 @@ out: return -EINVAL; } -static u32 esp4_get_max_size(struct xfrm_state *x, int mtu) +static u32 esp4_get_mtu(struct xfrm_state *x, int mtu) { struct esp_data *esp = x->data; u32 blksize = ALIGN(crypto_blkcipher_blocksize(esp->conf.tfm), 4); - int enclen = 0; + u32 align = max_t(u32, blksize, esp->conf.padlen); + u32 rem; + + mtu -= x->props.header_len + esp->auth.icv_trunc_len; + rem = mtu & (align - 1); + mtu &= ~(align - 1); switch (x->props.mode) { case XFRM_MODE_TUNNEL: - mtu = ALIGN(mtu +2, blksize); break; default: case XFRM_MODE_TRANSPORT: /* The worst case */ - mtu = ALIGN(mtu + 2, 4) + blksize - 4; + mtu -= blksize - 4; + mtu += min_t(u32, blksize - 4, rem); break; case XFRM_MODE_BEET: /* The worst case. */ - enclen = IPV4_BEET_PHMAXLEN; - mtu = ALIGN(mtu + enclen + 2, blksize); + mtu -= IPV4_BEET_PHMAXLEN; + mtu += min_t(u32, IPV4_BEET_PHMAXLEN, rem); break; } - if (esp->conf.padlen) - mtu = ALIGN(mtu, esp->conf.padlen); - - return mtu + x->props.header_len + esp->auth.icv_trunc_len - enclen; + return mtu - 2; } static void esp4_err(struct sk_buff *skb, u32 info) @@ -340,6 +342,7 @@ static int esp_init_state(struct xfrm_state *x) { struct esp_data *esp = NULL; struct crypto_blkcipher *tfm; + u32 align; /* null auth and encryption can have zero length keys */ if (x->aalg) { @@ -421,7 +424,10 @@ static int esp_init_state(struct xfrm_state *x) } } x->data = esp; - x->props.trailer_len = esp4_get_max_size(x, 0) - x->props.header_len; + align = ALIGN(crypto_blkcipher_blocksize(esp->conf.tfm), 4); + if (esp->conf.padlen) + align = max_t(u32, align, esp->conf.padlen); + x->props.trailer_len = align + 1 + esp->auth.icv_trunc_len; return 0; error: @@ -438,7 +444,7 @@ static struct xfrm_type esp_type = .proto = IPPROTO_ESP, .init_state = esp_init_state, .destructor = esp_destroy, - .get_max_size = esp4_get_max_size, + .get_mtu = esp4_get_mtu, .input = esp_input, .output = esp_output }; diff --git a/net/ipv6/esp6.c b/net/ipv6/esp6.c index b8e8914..61af22d 100644 --- a/net/ipv6/esp6.c +++ b/net/ipv6/esp6.c @@ -235,22 +235,24 @@ out: return ret; } -static u32 esp6_get_max_size(struct xfrm_state *x, int mtu) +static u32 esp6_get_mtu(struct xfrm_state *x, int mtu) { struct esp_data *esp = x->data; u32 blksize = ALIGN(crypto_blkcipher_blocksize(esp->conf.tfm), 4); + u32 align = max_t(u32, blksize, esp->conf.padlen); + u32 rem; - if (x->props.mode == XFRM_MODE_TUNNEL) { - mtu = ALIGN(mtu + 2, blksize); - } else { - /* The worst case. */ + mtu -= x->props.header_len + esp->auth.icv_trunc_len; + rem = mtu & (align - 1); + mtu &= ~(align - 1); + + if (x->props.mode != XFRM_MODE_TUNNEL) { u32 padsize = ((blksize - 1) & 7) + 1; - mtu = ALIGN(mtu + 2, padsize) + blksize - padsize; + mtu -= blksize - padsize; + mtu += min_t(u32, blksize - padsize, rem); } - if (esp->conf.padlen) - mtu = ALIGN(mtu, esp->conf.padlen); - return mtu + x->props.header_len + esp->auth.icv_trunc_len; + return mtu - 2; } static void esp6_err(struct sk_buff *skb, struct inet6_skb_parm *opt, @@ -380,7 +382,7 @@ static struct xfrm_type esp6_type = .proto = IPPROTO_ESP, .init_state = esp6_init_state, .destructor = esp6_destroy, - .get_max_size = esp6_get_max_size, + .get_mtu = esp6_get_mtu, .input = esp6_input, .output = esp6_output, .hdr_offset = xfrm6_find_1stfragopt, diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c index a237dfd..790b608 100644 --- a/net/xfrm/xfrm_state.c +++ b/net/xfrm/xfrm_state.c @@ -1666,37 +1666,17 @@ void xfrm_state_delete_tunnel(struct xfrm_state *x) } EXPORT_SYMBOL(xfrm_state_delete_tunnel); -/* - * This function is NOT optimal. For example, with ESP it will give an - * MTU that's usually two bytes short of being optimal. However, it will - * usually give an answer that's a multiple of 4 provided the input is - * also a multiple of 4. - */ int xfrm_state_mtu(struct xfrm_state *x, int mtu) { - int res = mtu; - - res -= x->props.header_len; - - for (;;) { - int m = res; - - if (m < 68) - return 68; - - spin_lock_bh(&x->lock); - if (x->km.state == XFRM_STATE_VALID && - x->type && x->type->get_max_size) - m = x->type->get_max_size(x, m); - else - m += x->props.header_len; - spin_unlock_bh(&x->lock); - - if (m <= mtu) - break; - res -= (m - mtu); - } + int res; + spin_lock_bh(&x->lock); + if (x->km.state == XFRM_STATE_VALID && + x->type && x->type->get_mtu) + res = x->type->get_mtu(x, mtu); + else + res = mtu; + spin_unlock_bh(&x->lock); return res; }