Revert rev 1.109:
| ixl/ice(4): use 128 segments for DMA maps of TSO packets
|
| This avoids unnecessary m_defrag() calls and gain some performance.
|
| with tweaks for kettenis and bluhm
|
| ok kettenis, bluhm
This seems to cause oactive stalls and ixl is too common to have it broken
like this. I prefer unnecessary m_defrag() calls over hanging interfaces.
use strtonum() instead of atoi(), and error out for bad numbers
This generates a host-order number, so the ntohs() for getservbyport()
was wrong, that should always have been htons(). The transform is
the same, but misleading.
ok tb
const correct X509_VERIFY_PARAM_get_hostflags()
This is currently an internal helper only used by a regress test.
We'll have to expose in the public API for Python 3.14:
https://github.com/libressl/portable/issues/1202
Add SMU support to amdpmc(4). The SMU is a microcontroller buried deep
in the bowels of AMD SoCs and needs to be tickled in order to reach the
lowest power states in suspend. It also provides some information on
whether we actually reached those lower states.
On most machines we will need to tell the USB4 controllers to go to sleep
to reach these states. So it is likely that we won't reach S0i3 yet on
those machines.
ok mlarkin@, deraadt@
Fix various errors on big-endian systems. With this ice(4) works on sparc64.
In ice_copy_rxq_ctx_to_hw() the buffer passed in is already in little
endian and so one needs to call bus_space_write_raw_4(). Wrap that into
ICE_WRITE_RAW() to be consistent.
In ice_txeof() the calculation of dtype was all messed up. The 64bit
descriptor word cmd_type_offset_bsz needs to be converted to host byte
order and then all the masking and shifting can happen. Also there is
no need to byte swap ICE_TX_DESC_DTYPE_DESC_DONE since dtype is already
in host byte order.
On code inspection another unneeded byte swap was found in
ice_tx_setup_offload(). The vlan tag needs to be shifted into position
unswapped and the full word is swapped at the end.
OK stsp@
Remove unused sequence member from x509_revoked_st
To allow binary search for looking up if a cert was revoked in a CRL,
the list of revoked serial numbers is sorted in crl_lookup(). On the
other hand, to be able to output the DER that was actually signed by
the issuer, the original order needs to be remembered.
Before the encoding was cached, there was a mechanism that would restore
the original order on serialization using the .sequence member. This was
done without a lock and was thus racy (hilarity would ensue if one thread
performed a CRL lookup while another thread serialized the same CRL). When
the racy mechanism was removed in 2004, the only reader of .sequence,
X509_REVOKED_seq_cmp(), was also removed, and this piece of dead code was
left behind. Garbage collect it.
ok kenjiro
Improve JH7110 support:
- Round to the nearest achievable clock rate instead of rounding down.
- Make sure we don't set a divider to zero.
- Fully initialize PLL0 when setting its rate.
- Bump PLL0 rate regardless of what the firmware configures it to.
This avoids issues with firmware based on upstream U-Boot.
ok jca@
Clarify CCR file format: the payload is encapsulated inside an OCTET STRING
Brought up by William McCall
This is a mechanical change.
discussed with / OK tb@
When tab-completing a filename, ensure that the completed string
does not end up mid-way through a multibyte character, as this
will cause a fatal() later on.
based on GHPR#587 from @TaoistBrickscarrier; feedback tb@ kevlo@
ok dtucker@
Fix association to access points which have all 802.11b rates disabled.
When I replaced ieee80211_iserp_sta() with a check for the extended
rates information element I wrongly assumed that this information
element would always be present if 11g is supported. However, the
xrates IE only appears in 11b/g mixed mode. APs which have 802.11b
rates completely disabled can announce all rates via the regular
rates IE and never send the xrates IE.
Ensure that we recognize such APs as 11g-capable regardless of the
missing xrates IE. Otherwise assocation can fail due to a mismatch
of basic rates.
Problem reported by landry@
ok phessler@ sthen@
implement "checksum offload" between rport pairs.
this allows the kernel to skip ip/tcp/udp checksum calculation for
packets between rdomains.
rport interfaces advertise checksum offload capabilities to the
stack, so the stack will set CSUM_OUT mbuf flags rather than do the
checksum calculation for packets being transmitted on an port
interface. when these packets are sent back into the stack by the
partner rport interface, the CSUM_OUT flags get mapped to equivalent
CSUM_IN_OK flags. this is modelled on how lo(4) does the same thing.