Re: mcl2k2 mbuf clusters

[Mike Belopuhov]

On 14 October 2016 at 12:40, Mike Belopuhov  wrote:
> On Fri, Oct 14, 2016 at 07:58 +0200, Claudio Jeker wrote:
>> It is time to put the nasty comment from rl(4) into em(4) and ix(4).
>> Everybody knew how bad realtek was but thinks Intel nics are good. The
>> truth is that modern Intel nic are as bad as the cheepest and crapiest
>> 10/100 Mbps Ethernet chips from the last millenium.
>>
>> --
>> :wq Claudio
>>
>
> Then don't use them.  Vanity won't fix anything (if there's anything
> to fix at all).

I don't mean to offend you personally.  It's directed at all of us,
me included.  I apologize if that sounded too harsh.

Re: mcl2k2 mbuf clusters

[Mike Belopuhov]

On Fri, Oct 14, 2016 at 15:48 +1000, David Gwynne wrote:
> this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
> can be used on some very common nics that have annoying constraints
> on their rx descriptors.
> 
> this in turn simplifies the code in those drivers and lets them
> always operate on ETHER_ALIGN boundaries.
>

I would agree for using it under the _STRICT_ALIGNED ifdef instead
of wasting a page, but not in the generic case.  No technical reason
to do it.  Code simplifications brought by this change are miniscule.
In fact it adds more code to the uipc_mbuf than saves in the driver.

> the pool is cheap, pages will only be allocated in it if something
> asks for them, and it keeps this complexity out of the drivers.
> 
> ok?
> 

Not OK.  Not w/o *at least* the performance assessment of this change.
Right now it's just a distraction from more important issues.

Re: mcl2k2 mbuf clusters

[Mike Belopuhov]

On Fri, Oct 14, 2016 at 07:58 +0200, Claudio Jeker wrote:
> It is time to put the nasty comment from rl(4) into em(4) and ix(4).
> Everybody knew how bad realtek was but thinks Intel nics are good. The
> truth is that modern Intel nic are as bad as the cheepest and crapiest
> 10/100 Mbps Ethernet chips from the last millenium.
> 
> -- 
> :wq Claudio
> 

Then don't use them.  Vanity won't fix anything (if there's anything
to fix at all).

Re: mcl2k2 mbuf clusters

[David Gwynne]

> On 14 Oct 2016, at 16:17, Ted Unangst  wrote:
> 
> David Gwynne wrote:
>> this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
>> can be used on some very common nics that have annoying constraints
>> on their rx descriptors.
>> 
>> this in turn simplifies the code in those drivers and lets them
>> always operate on ETHER_ALIGN boundaries.
>> 
>> the pool is cheap, pages will only be allocated in it if something
>> asks for them, and it keeps this complexity out of the drivers.
> 
> pool is efficient, but it can't work miracles. on archs like amd64, this is
> going to only pack 15 clusters in a 32k "page" instead of 16.

it is only about 70 bytes per cluster though. i could raise the size by that to 
avoid slack space if you want, or we can let the page colouring have some fun 
here.

> 
> if i understand correctly, currently em is using 4k clusters on strict
> alignment? i think it makes a lot of sense to add a 2k2 pool because that
> means using less memory. but not strict alignment shouldn't be forced to use
> more memory. that's a regression.

you could also argue that 2k clusters waste ~500 bytes in the most common case 
since the vast majority of large packets are still only 1514 bytes. we could 
pack 10 1600 byte clusters into 16k instead of the 8 we do now.

more seriously though, im happy to burn a bit more memory to keep the code 
simple, and overall our memory consumption by network cards is much lower than 
other platforms because of the rx ring moderation we do.

dlg

> 
> 
>> 
>> ok?
>> 
>> Index: net/if.h
>> ===
>> RCS file: /cvs/src/sys/net/if.h,v
>> retrieving revision 1.179
>> diff -u -p -r1.179 if.h
>> --- net/if.h 4 Sep 2016 15:10:59 -   1.179
>> +++ net/if.h 14 Oct 2016 03:46:22 -
>> @@ -68,7 +68,7 @@ struct if_clonereq {
>>  char*ifcr_buffer;   /* buffer for cloner names */
>> };
>> 
>> -#define MCLPOOLS7   /* number of cluster pools */
>> +#define MCLPOOLS8   /* number of cluster pools */
>> 
>> struct if_rxring {
>>  int rxr_adjusted;
>> Index: kern/uipc_mbuf.c
>> ===
>> RCS file: /cvs/src/sys/kern/uipc_mbuf.c,v
>> retrieving revision 1.233
>> diff -u -p -r1.233 uipc_mbuf.c
>> --- kern/uipc_mbuf.c 10 Oct 2016 00:41:17 -  1.233
>> +++ kern/uipc_mbuf.c 14 Oct 2016 03:46:22 -
>> @@ -107,6 +107,7 @@ struct   pool mtagpool;
>> /* mbuf cluster pools */
>> u_intmclsizes[MCLPOOLS] = {
>>  MCLBYTES,   /* must be at slot 0 */
>> +MCLBYTES + 2,   /* ETHER_ALIGNED 2k mbufs */
>>  4 * 1024,
>>  8 * 1024,
>>  9 * 1024,
>> @@ -142,6 +143,7 @@ void
>> mbinit(void)
>> {
>>  int i;
>> +unsigned int lowbits;
>> 
>> #if DIAGNOSTIC
>>  if (mclsizes[0] != MCLBYTES)
>> @@ -158,9 +160,15 @@ mbinit(void)
>>  IPL_NET, 0, "mtagpl", NULL);
>> 
>>  for (i = 0; i < nitems(mclsizes); i++) {
>> -snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
>> -mclsizes[i] >> 10);
>> -pool_init([i], mclsizes[i], 0, IPL_NET, 0,
>> +lowbits = mclsizes[i] & ((1 << 10) - 1);
>> +if (lowbits) {
>> +snprintf(mclnames[i], sizeof(mclnames[0]),
>> +"mcl%dk%u", mclsizes[i] >> 10, lowbits);
>> +} else {
>> +snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
>> +mclsizes[i] >> 10);
>> +}
>> +pool_init([i], mclsizes[i], 64, IPL_NET, 0,
>>  mclnames[i], NULL);
>>  pool_set_constraints([i], _dma_contig);
>>  pool_setlowat([i], mcllowat);
>> Index: dev/pci/if_em.c
>> ===
>> RCS file: /cvs/src/sys/dev/pci/if_em.c,v
>> retrieving revision 1.331
>> diff -u -p -r1.331 if_em.c
>> --- dev/pci/if_em.c  13 Apr 2016 10:34:32 -  1.331
>> +++ dev/pci/if_em.c  14 Oct 2016 03:47:21 -
>> @@ -2450,9 +2450,7 @@ em_get_buf(struct em_softc *sc, int i)
>>  return (ENOBUFS);
>>  }
>>  m->m_len = m->m_pkthdr.len = EM_MCLBYTES;
>> -#ifdef __STRICT_ALIGNMENT
>>  m_adj(m, ETHER_ALIGN);
>> -#endif
>> 
>>  error = bus_dmamap_load_mbuf(sc->sc_dmat, pkt->pkt_map,
>>  m, BUS_DMA_NOWAIT);
>> Index: dev/pci/if_ix.c
>> ===
>> RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
>> retrieving revision 1.132
>> diff -u -p -r1.132 if_ix.c
>> --- dev/pci/if_ix.c  13 Apr 2016 10:34:32 -  1.132
>> +++ dev/pci/if_ix.c  14 Oct 2016 03:47:21 -
>> @@ -616,11 +616,7 @@ ixgbe_init(void *arg)
>>  ixgbe_initialize_transmit_units(sc);
>> 
>>  /* Use 2k clusters, even for jumbo frames */
>> -#ifdef __STRICT_ALIGNMENT
>>  sc->rx_mbuf_sz = MCLBYTES + ETHER_ALIGN;
>> -#else
>> -   

Re: mcl2k2 mbuf clusters

[Ted Unangst]

David Gwynne wrote:
> this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
> can be used on some very common nics that have annoying constraints
> on their rx descriptors.
> 
> this in turn simplifies the code in those drivers and lets them
> always operate on ETHER_ALIGN boundaries.
> 
> the pool is cheap, pages will only be allocated in it if something
> asks for them, and it keeps this complexity out of the drivers.

pool is efficient, but it can't work miracles. on archs like amd64, this is
going to only pack 15 clusters in a 32k "page" instead of 16.

if i understand correctly, currently em is using 4k clusters on strict
alignment? i think it makes a lot of sense to add a 2k2 pool because that
means using less memory. but not strict alignment shouldn't be forced to use
more memory. that's a regression.


> 
> ok?
> 
> Index: net/if.h
> ===
> RCS file: /cvs/src/sys/net/if.h,v
> retrieving revision 1.179
> diff -u -p -r1.179 if.h
> --- net/if.h  4 Sep 2016 15:10:59 -   1.179
> +++ net/if.h  14 Oct 2016 03:46:22 -
> @@ -68,7 +68,7 @@ struct if_clonereq {
>   char*ifcr_buffer;   /* buffer for cloner names */
>  };
>  
> -#define MCLPOOLS 7   /* number of cluster pools */
> +#define MCLPOOLS 8   /* number of cluster pools */
>  
>  struct if_rxring {
>   int rxr_adjusted;
> Index: kern/uipc_mbuf.c
> ===
> RCS file: /cvs/src/sys/kern/uipc_mbuf.c,v
> retrieving revision 1.233
> diff -u -p -r1.233 uipc_mbuf.c
> --- kern/uipc_mbuf.c  10 Oct 2016 00:41:17 -  1.233
> +++ kern/uipc_mbuf.c  14 Oct 2016 03:46:22 -
> @@ -107,6 +107,7 @@ structpool mtagpool;
>  /* mbuf cluster pools */
>  u_intmclsizes[MCLPOOLS] = {
>   MCLBYTES,   /* must be at slot 0 */
> + MCLBYTES + 2,   /* ETHER_ALIGNED 2k mbufs */
>   4 * 1024,
>   8 * 1024,
>   9 * 1024,
> @@ -142,6 +143,7 @@ void
>  mbinit(void)
>  {
>   int i;
> + unsigned int lowbits;
>  
>  #if DIAGNOSTIC
>   if (mclsizes[0] != MCLBYTES)
> @@ -158,9 +160,15 @@ mbinit(void)
>   IPL_NET, 0, "mtagpl", NULL);
>  
>   for (i = 0; i < nitems(mclsizes); i++) {
> - snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
> - mclsizes[i] >> 10);
> - pool_init([i], mclsizes[i], 0, IPL_NET, 0,
> + lowbits = mclsizes[i] & ((1 << 10) - 1);
> + if (lowbits) {
> + snprintf(mclnames[i], sizeof(mclnames[0]),
> + "mcl%dk%u", mclsizes[i] >> 10, lowbits);
> + } else {
> + snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
> + mclsizes[i] >> 10);
> + }
> + pool_init([i], mclsizes[i], 64, IPL_NET, 0,
>   mclnames[i], NULL);
>   pool_set_constraints([i], _dma_contig);
>   pool_setlowat([i], mcllowat);
> Index: dev/pci/if_em.c
> ===
> RCS file: /cvs/src/sys/dev/pci/if_em.c,v
> retrieving revision 1.331
> diff -u -p -r1.331 if_em.c
> --- dev/pci/if_em.c   13 Apr 2016 10:34:32 -  1.331
> +++ dev/pci/if_em.c   14 Oct 2016 03:47:21 -
> @@ -2450,9 +2450,7 @@ em_get_buf(struct em_softc *sc, int i)
>   return (ENOBUFS);
>   }
>   m->m_len = m->m_pkthdr.len = EM_MCLBYTES;
> -#ifdef __STRICT_ALIGNMENT
>   m_adj(m, ETHER_ALIGN);
> -#endif
>  
>   error = bus_dmamap_load_mbuf(sc->sc_dmat, pkt->pkt_map,
>   m, BUS_DMA_NOWAIT);
> Index: dev/pci/if_ix.c
> ===
> RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
> retrieving revision 1.132
> diff -u -p -r1.132 if_ix.c
> --- dev/pci/if_ix.c   13 Apr 2016 10:34:32 -  1.132
> +++ dev/pci/if_ix.c   14 Oct 2016 03:47:21 -
> @@ -616,11 +616,7 @@ ixgbe_init(void *arg)
>   ixgbe_initialize_transmit_units(sc);
>  
>   /* Use 2k clusters, even for jumbo frames */
> -#ifdef __STRICT_ALIGNMENT
>   sc->rx_mbuf_sz = MCLBYTES + ETHER_ALIGN;
> -#else
> - sc->rx_mbuf_sz = MCLBYTES;
> -#endif
>  
>   /* Prepare receive descriptors and buffers */
>   if (ixgbe_setup_receive_structures(sc)) {
> @@ -2458,9 +2454,7 @@ ixgbe_get_buf(struct rx_ring *rxr, int i
>   return (ENOBUFS);
>  
>   mp->m_len = mp->m_pkthdr.len = sc->rx_mbuf_sz;
> -#ifdef __STRICT_ALIGNMENT
>   m_adj(mp, ETHER_ALIGN);
> -#endif
>  
>   error = bus_dmamap_load_mbuf(rxr->rxdma.dma_tag, rxbuf->map,
>   mp, BUS_DMA_NOWAIT);
> @@ -2667,11 +2661,7 @@ ixgbe_initialize_receive_units(struct ix
>   hlreg |= IXGBE_HLREG0_JUMBOEN;
>   IXGBE_WRITE_REG(>hw, IXGBE_HLREG0, hlreg);
>  
> -#ifdef __STRICT_ALIGNMENT
>   bufsz = (sc->rx_mbuf_sz - ETHER_ALIGN) >> 

Re: mcl2k2 mbuf clusters

[Claudio Jeker]

It is time to put the nasty comment from rl(4) into em(4) and ix(4).
Everybody knew how bad realtek was but thinks Intel nics are good. The
truth is that modern Intel nic are as bad as the cheepest and crapiest
10/100 Mbps Ethernet chips from the last millenium.

-- 
:wq Claudio

On Fri, Oct 14, 2016 at 03:48:18PM +1000, David Gwynne wrote:
> this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
> can be used on some very common nics that have annoying constraints
> on their rx descriptors.
> 
> this in turn simplifies the code in those drivers and lets them
> always operate on ETHER_ALIGN boundaries.
> 
> the pool is cheap, pages will only be allocated in it if something
> asks for them, and it keeps this complexity out of the drivers.
> 
> ok?
> 
> Index: net/if.h
> ===
> RCS file: /cvs/src/sys/net/if.h,v
> retrieving revision 1.179
> diff -u -p -r1.179 if.h
> --- net/if.h  4 Sep 2016 15:10:59 -   1.179
> +++ net/if.h  14 Oct 2016 03:46:22 -
> @@ -68,7 +68,7 @@ struct if_clonereq {
>   char*ifcr_buffer;   /* buffer for cloner names */
>  };
>  
> -#define MCLPOOLS 7   /* number of cluster pools */
> +#define MCLPOOLS 8   /* number of cluster pools */
>  
>  struct if_rxring {
>   int rxr_adjusted;
> Index: kern/uipc_mbuf.c
> ===
> RCS file: /cvs/src/sys/kern/uipc_mbuf.c,v
> retrieving revision 1.233
> diff -u -p -r1.233 uipc_mbuf.c
> --- kern/uipc_mbuf.c  10 Oct 2016 00:41:17 -  1.233
> +++ kern/uipc_mbuf.c  14 Oct 2016 03:46:22 -
> @@ -107,6 +107,7 @@ structpool mtagpool;
>  /* mbuf cluster pools */
>  u_intmclsizes[MCLPOOLS] = {
>   MCLBYTES,   /* must be at slot 0 */
> + MCLBYTES + 2,   /* ETHER_ALIGNED 2k mbufs */
>   4 * 1024,
>   8 * 1024,
>   9 * 1024,
> @@ -142,6 +143,7 @@ void
>  mbinit(void)
>  {
>   int i;
> + unsigned int lowbits;
>  
>  #if DIAGNOSTIC
>   if (mclsizes[0] != MCLBYTES)
> @@ -158,9 +160,15 @@ mbinit(void)
>   IPL_NET, 0, "mtagpl", NULL);
>  
>   for (i = 0; i < nitems(mclsizes); i++) {
> - snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
> - mclsizes[i] >> 10);
> - pool_init([i], mclsizes[i], 0, IPL_NET, 0,
> + lowbits = mclsizes[i] & ((1 << 10) - 1);
> + if (lowbits) {
> + snprintf(mclnames[i], sizeof(mclnames[0]),
> + "mcl%dk%u", mclsizes[i] >> 10, lowbits);
> + } else {
> + snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
> + mclsizes[i] >> 10);
> + }
> + pool_init([i], mclsizes[i], 64, IPL_NET, 0,
>   mclnames[i], NULL);
>   pool_set_constraints([i], _dma_contig);
>   pool_setlowat([i], mcllowat);
> Index: dev/pci/if_em.c
> ===
> RCS file: /cvs/src/sys/dev/pci/if_em.c,v
> retrieving revision 1.331
> diff -u -p -r1.331 if_em.c
> --- dev/pci/if_em.c   13 Apr 2016 10:34:32 -  1.331
> +++ dev/pci/if_em.c   14 Oct 2016 03:47:21 -
> @@ -2450,9 +2450,7 @@ em_get_buf(struct em_softc *sc, int i)
>   return (ENOBUFS);
>   }
>   m->m_len = m->m_pkthdr.len = EM_MCLBYTES;
> -#ifdef __STRICT_ALIGNMENT
>   m_adj(m, ETHER_ALIGN);
> -#endif
>  
>   error = bus_dmamap_load_mbuf(sc->sc_dmat, pkt->pkt_map,
>   m, BUS_DMA_NOWAIT);
> Index: dev/pci/if_ix.c
> ===
> RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
> retrieving revision 1.132
> diff -u -p -r1.132 if_ix.c
> --- dev/pci/if_ix.c   13 Apr 2016 10:34:32 -  1.132
> +++ dev/pci/if_ix.c   14 Oct 2016 03:47:21 -
> @@ -616,11 +616,7 @@ ixgbe_init(void *arg)
>   ixgbe_initialize_transmit_units(sc);
>  
>   /* Use 2k clusters, even for jumbo frames */
> -#ifdef __STRICT_ALIGNMENT
>   sc->rx_mbuf_sz = MCLBYTES + ETHER_ALIGN;
> -#else
> - sc->rx_mbuf_sz = MCLBYTES;
> -#endif
>  
>   /* Prepare receive descriptors and buffers */
>   if (ixgbe_setup_receive_structures(sc)) {
> @@ -2458,9 +2454,7 @@ ixgbe_get_buf(struct rx_ring *rxr, int i
>   return (ENOBUFS);
>  
>   mp->m_len = mp->m_pkthdr.len = sc->rx_mbuf_sz;
> -#ifdef __STRICT_ALIGNMENT
>   m_adj(mp, ETHER_ALIGN);
> -#endif
>  
>   error = bus_dmamap_load_mbuf(rxr->rxdma.dma_tag, rxbuf->map,
>   mp, BUS_DMA_NOWAIT);
> @@ -2667,11 +2661,7 @@ ixgbe_initialize_receive_units(struct ix
>   hlreg |= IXGBE_HLREG0_JUMBOEN;
>   IXGBE_WRITE_REG(>hw, IXGBE_HLREG0, hlreg);
>  
> -#ifdef __STRICT_ALIGNMENT
>   bufsz = (sc->rx_mbuf_sz - ETHER_ALIGN) >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;
> -#else
> - bufsz = sc->rx_mbuf_sz >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;

mcl2k2 mbuf clusters

[David Gwynne]

this adds a pool backend for MCLGETI thats 2k+2 bytes in size, which
can be used on some very common nics that have annoying constraints
on their rx descriptors.

this in turn simplifies the code in those drivers and lets them
always operate on ETHER_ALIGN boundaries.

the pool is cheap, pages will only be allocated in it if something
asks for them, and it keeps this complexity out of the drivers.

ok?

Index: net/if.h
===
RCS file: /cvs/src/sys/net/if.h,v
retrieving revision 1.179
diff -u -p -r1.179 if.h
--- net/if.h4 Sep 2016 15:10:59 -   1.179
+++ net/if.h14 Oct 2016 03:46:22 -
@@ -68,7 +68,7 @@ struct if_clonereq {
char*ifcr_buffer;   /* buffer for cloner names */
 };
 
-#define MCLPOOLS   7   /* number of cluster pools */
+#define MCLPOOLS   8   /* number of cluster pools */
 
 struct if_rxring {
int rxr_adjusted;
Index: kern/uipc_mbuf.c
===
RCS file: /cvs/src/sys/kern/uipc_mbuf.c,v
retrieving revision 1.233
diff -u -p -r1.233 uipc_mbuf.c
--- kern/uipc_mbuf.c10 Oct 2016 00:41:17 -  1.233
+++ kern/uipc_mbuf.c14 Oct 2016 03:46:22 -
@@ -107,6 +107,7 @@ struct  pool mtagpool;
 /* mbuf cluster pools */
 u_int  mclsizes[MCLPOOLS] = {
MCLBYTES,   /* must be at slot 0 */
+   MCLBYTES + 2,   /* ETHER_ALIGNED 2k mbufs */
4 * 1024,
8 * 1024,
9 * 1024,
@@ -142,6 +143,7 @@ void
 mbinit(void)
 {
int i;
+   unsigned int lowbits;
 
 #if DIAGNOSTIC
if (mclsizes[0] != MCLBYTES)
@@ -158,9 +160,15 @@ mbinit(void)
IPL_NET, 0, "mtagpl", NULL);
 
for (i = 0; i < nitems(mclsizes); i++) {
-   snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
-   mclsizes[i] >> 10);
-   pool_init([i], mclsizes[i], 0, IPL_NET, 0,
+   lowbits = mclsizes[i] & ((1 << 10) - 1);
+   if (lowbits) {
+   snprintf(mclnames[i], sizeof(mclnames[0]),
+   "mcl%dk%u", mclsizes[i] >> 10, lowbits);
+   } else {
+   snprintf(mclnames[i], sizeof(mclnames[0]), "mcl%dk",
+   mclsizes[i] >> 10);
+   }
+   pool_init([i], mclsizes[i], 64, IPL_NET, 0,
mclnames[i], NULL);
pool_set_constraints([i], _dma_contig);
pool_setlowat([i], mcllowat);
Index: dev/pci/if_em.c
===
RCS file: /cvs/src/sys/dev/pci/if_em.c,v
retrieving revision 1.331
diff -u -p -r1.331 if_em.c
--- dev/pci/if_em.c 13 Apr 2016 10:34:32 -  1.331
+++ dev/pci/if_em.c 14 Oct 2016 03:47:21 -
@@ -2450,9 +2450,7 @@ em_get_buf(struct em_softc *sc, int i)
return (ENOBUFS);
}
m->m_len = m->m_pkthdr.len = EM_MCLBYTES;
-#ifdef __STRICT_ALIGNMENT
m_adj(m, ETHER_ALIGN);
-#endif
 
error = bus_dmamap_load_mbuf(sc->sc_dmat, pkt->pkt_map,
m, BUS_DMA_NOWAIT);
Index: dev/pci/if_ix.c
===
RCS file: /cvs/src/sys/dev/pci/if_ix.c,v
retrieving revision 1.132
diff -u -p -r1.132 if_ix.c
--- dev/pci/if_ix.c 13 Apr 2016 10:34:32 -  1.132
+++ dev/pci/if_ix.c 14 Oct 2016 03:47:21 -
@@ -616,11 +616,7 @@ ixgbe_init(void *arg)
ixgbe_initialize_transmit_units(sc);
 
/* Use 2k clusters, even for jumbo frames */
-#ifdef __STRICT_ALIGNMENT
sc->rx_mbuf_sz = MCLBYTES + ETHER_ALIGN;
-#else
-   sc->rx_mbuf_sz = MCLBYTES;
-#endif
 
/* Prepare receive descriptors and buffers */
if (ixgbe_setup_receive_structures(sc)) {
@@ -2458,9 +2454,7 @@ ixgbe_get_buf(struct rx_ring *rxr, int i
return (ENOBUFS);
 
mp->m_len = mp->m_pkthdr.len = sc->rx_mbuf_sz;
-#ifdef __STRICT_ALIGNMENT
m_adj(mp, ETHER_ALIGN);
-#endif
 
error = bus_dmamap_load_mbuf(rxr->rxdma.dma_tag, rxbuf->map,
mp, BUS_DMA_NOWAIT);
@@ -2667,11 +2661,7 @@ ixgbe_initialize_receive_units(struct ix
hlreg |= IXGBE_HLREG0_JUMBOEN;
IXGBE_WRITE_REG(>hw, IXGBE_HLREG0, hlreg);
 
-#ifdef __STRICT_ALIGNMENT
bufsz = (sc->rx_mbuf_sz - ETHER_ALIGN) >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;
-#else
-   bufsz = sc->rx_mbuf_sz >> IXGBE_SRRCTL_BSIZEPKT_SHIFT;
-#endif
 
for (i = 0; i < sc->num_queues; i++, rxr++) {
uint64_t rdba = rxr->rxdma.dma_map->dm_segs[0].ds_addr;
Index: dev/pci/if_em.h
===
RCS file: /cvs/src/sys/dev/pci/if_em.h,v
retrieving revision 1.72
diff -u -p -r1.72 if_em.h
--- dev/pci/if_em.h 18 Feb 2016 14:24:39 -  1.72
+++ dev/pci/if_em.h 14 Oct 2016 03:48:32 -
@@ -263,11 +263,7 @@ typedef int