Emmanuel Vadot [Sat, 1 Dec 2018 20:26:59 +0000 (20:26 +0000)]
arm64: rockchip: Add RK3399_CLK_PLL
PLLs on the RK3399 are different than the ones on the RK3328.
Add a new type and some dedicated recalc and set_freq functions.
Rename the RK3328 dedicated rk_clk_pll function with rk3328_ prefix.
Cy Schubert [Sat, 1 Dec 2018 17:59:41 +0000 (17:59 +0000)]
Restore handling of PMTU discovery, removed through an unifdef(1)
following the MFV of r254219 into r255332. In addition the 'FreeBSD'
macro was never defined in ipfilter 5.1.2 thus it never would have
been enabled in the first place.
This work is prompted by a general cleanup of the IP Filter code
prompted by working to resolve a PR. More to follow.
Allow to create swap zone larger than v_page_count / 2.
If user configured the maxswapzone tunable, just take the literal
value for the initial zone sizing attempt. Before, it was only
possible to reduce the zone by the tunable.
While there, correct the message which was not correct when zone
creation rounded the size up.
Reported by: jmg
Reviewed by: markj
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D18381
Kristof Provost [Sat, 1 Dec 2018 09:58:21 +0000 (09:58 +0000)]
pf: Fix panic on overlapping interface names
In rare situations[*] it's possible for two different interfaces to have
the same name. This confuses pf, because kifs are indexed by name (which
is assumed to be unique). As a result we can end up trying to
if_rele(NULL), which panics.
Explicitly checking the ifp pointer before if_rele() prevents the panic.
Note pf will likely behave in unexpected ways on the the overlapping
interfaces.
[*] Insert an interface in a vnet jail. Rename it to an interface which
exists on the host. Remove the jail. There are now two interfaces with
the same name in the host.
Adapt the fix in r341008 to correctly work with EBR.
IFNET_RLOCK_NOSLEEP() is epoch_enter_preempt() in FreeBSD 12+. Holding
it in sysctl_rtsock() doesn't protect us from ifnet unlinking, because
unlinking occurs with IFNET_WLOCK(), that is rw_wlock+sx_xlock, and it
doesn check that concurrent code is running in epoch section. But while
we are in epoch section, we should be able to do access to ifnet's
fields, even it was unlinked. Thus do not change if_addr and if_hw_addr
fields in ifnet_detach_internal() to NULL, since rtsock code can do
access to these fields and this is allowed while it is running in epoch
section.
This should fix the race, when ifnet_detach_internal() unlinks ifnet
after we checked it for IFF_DYING in sysctl_dumpentry.
Move free(ifp->if_hw_addr) into ifnet_free_internal(). Also remove the
NULL check for ifp->if_description, since free(9) can correctly handle
NULL pointer.
Emmanuel Vadot [Fri, 30 Nov 2018 10:31:30 +0000 (10:31 +0000)]
arm64: allwinner: Add 792Mhz frequency to sun50i-a64-opp
This is the frequency of the cpu on the Pinebook so add it to make
cpufreq find the current setting.
Note that this dtbo on the Pinebook doesn't work right now as u-boot
dtb doesn't have symbols and so it fails to apply. Linux 4.20 have
the dts and will be imported once taggued.
Toomas Soome [Fri, 30 Nov 2018 08:42:14 +0000 (08:42 +0000)]
loader.efi: fix EFI getchar() for multiple consoles
This fix is ported from illumos (issue #9970), the analysis and initial
implementation was done by John Levon.
See also: https://www.illumos.org/issues/9970
Currently, efi_cons_getchar() will wait for a key. While this seems to make
sense, the implementation of getchar() in common/console.c will loop across
getchar() for all consoles without doing ischar() first.
This means that if we've configured multiple consoles, we can't input into
the serial, as getchar() will be sat waiting for input only from efi_console.c
This patch does implement a bit more generic key buffer to support
translation of input keys, and we use generic efi_readkey() to reduce
duplication from calls from getchar() and poll().
Toomas Soome [Fri, 30 Nov 2018 08:01:11 +0000 (08:01 +0000)]
loader: create separate lists for fd, cd and hd, merge bioscd with biosdisk
Create unified block IO implementation in BIOS version, like it is done in UEFI
side. Implement fd, disk and cd device lists, this will split floppy devices
from disks and will allow us to have consistent, predictable device naming
(modulo BIOS issues).
Andrew Rybchenko [Fri, 30 Nov 2018 07:11:05 +0000 (07:11 +0000)]
sfxge(4): rollback last seen VLAN TCI if Tx packet is dropped
Early processing of a packet on transmit may change last seen
VLAN TCI in the queue context. If such a packet is eventually
dropped, last seen VLAN TCI must be set to its previous value.
Submitted by: Ivan Malov <Ivan.Malov at oktetlabs.ru>
Sponsored by: Solarflare Communications, Inc.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D18288
Andrew Rybchenko [Fri, 30 Nov 2018 07:10:32 +0000 (07:10 +0000)]
sfxge(4): update external port number calculation
Revise the external port calculation to support all
X2 port modes. The previous algorithm could not
handle different port numbering schemes on each cage.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18285
Andrew Rybchenko [Fri, 30 Nov 2018 07:10:20 +0000 (07:10 +0000)]
sfxge(4): correct annotations where NULL input is OK
Correct annotations where NULL input can be permitted
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18284
Andrew Rybchenko [Fri, 30 Nov 2018 07:10:09 +0000 (07:10 +0000)]
sfxge(4): support new link modes in the driver
Submitted by: Andy Moreton <amoreton at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18283
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:58 +0000 (07:09 +0000)]
sfxge(4): use transceiver ID when reading info
In efx_mcdi_phy_module_get_info() probe the
transceiver identification byte rather than assume
the module matches the fixed port type. This
supports scenarios such as a SFP mounted in a QSFP
port via a QSA module.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18282
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:46 +0000 (07:09 +0000)]
sfxge(4): add accessor to whole link status
Add a function which makes an MCDI GET_LINK request and
packages up the results. Currently, the get-link function
is triggered from several entry points which then pass
on or store selected parts of the data. When the driver
needs to obtain the current link state, it is more
efficient to do this in a single call.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18281
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:34 +0000 (07:09 +0000)]
sfxge(4): guard Rx scale code with corresponding option
Previously only some of the code was guarded by this which caused
a build error when EFSYS_OPT_RX_SCALE is 0 (e.g. in manftest).
Submitted by: Tom Millington <tmillington at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18280
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:23 +0000 (07:09 +0000)]
sfxge(4): infer port mode bandwidth from max link speed
Limit the port mode bandwidth calculations by the maximum
reported link speed. This system detects 25G vs 10G cards,
and 100G port modes vs 40G.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18279
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:11 +0000 (07:09 +0000)]
sfxge(4): support improvements to bandwidth calculations
Change the interface to ef10_nic_get_port_mode_bandwidth()
so more NIC information can be used to infer bandwidth
requirements. Huntington calculations separated out
completely.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18278
Andrew Rybchenko [Fri, 30 Nov 2018 07:09:00 +0000 (07:09 +0000)]
sfxge(4): add X2 port modes to bandwidth calculator
Add cases for the new port modes supported by X2 NICs.
Lane bandwidth is calculated for pre-X2 cards so is an
underestimate for X2 in 25G/100G modes.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18277
Andrew Rybchenko [Fri, 30 Nov 2018 07:08:50 +0000 (07:08 +0000)]
sfxge(4): update to current port mode terminology
>From Medford onwards, the newer constants enumerating
port modes should be used.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18276
Andrew Rybchenko [Fri, 30 Nov 2018 07:08:38 +0000 (07:08 +0000)]
sfxge(4): adjust PHY module info interface
Adjust data types in interface to permit the complete
module information buffer to be obtained in a single
call.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18275
Rearrange so the valid addresses are visible to the caller.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18274
Andrew Rybchenko [Fri, 30 Nov 2018 07:07:43 +0000 (07:07 +0000)]
sfxge(4): modify phy caps to indicate FEC request
The capability bits to request FEC modes are implicitly valid
when the corresponding FEC mode is a supported capability.
Drivers expect that it is only valid to advertise those
capabilities explicitly marked as supported. The capabilities
reported by firmware is modified with the implicit capabilities
to present the explicit model to drivers.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18270
Andrew Rybchenko [Fri, 30 Nov 2018 07:07:31 +0000 (07:07 +0000)]
sfxge(4): improve handling of legacy RSS hash flags
Client drivers may use either legacy flags, for example,
EFX_RX_HASH_TCPIPV4, or generalised flags, for example,
EFX_RX_HASH(IPV4_TCP, 4TUPLE), to configure RSS hash.
The libefx is able to recognise what scheme is used.
Legacy flags may be consumed directly by a chip-specific handler to
configure the NIC, that is, on EF10, these flags can be used to fill
in legacy RSS mode field in MCDI request. Generalised flags can also
be directly used in EF10-specific handler as they are fully compatible
with additional fields of the same MCDI request.
Legacy flags undergo conversion to generalised flags before they
are consumed by a chip-specific handler. This conversion is used to
make sure that chip-specific handlers expect only generalised flags
in the input for the sake of clarity of the code.
Depending on firmware capabilities, a chip-specififc handler either
supplies the input to the NIC directly, for example,
EFX_RX_HASH(IPV4_TCP, 4TUPLE) flag will enable 4 bits in
RSS_CONTEXT_SET_FLAGS_IN_TCP_IPV4_RSS_MODE field on EF10, or takes
the opportunity to translate the input to enable bits which don't map
to the generic flag, like setting
RSS_CONTEXT_SET_FLAGS_IN_TOEPLITZ_TCPV4_EN on EF10 when the firmware
claims no support for additional modes.
However, this approach has introduced a severe problem which can be
reproduced with ultra-low-latency firmware variant. In order to enable
IP hash, EF10-specific handler requires the user to request 2-tuple
hash for IP-other, TCP and UDP traffic classes, unconditionally.
In example, IPv4 hash can be enabled using the following input:
EFX_RX_HASH(IPV4_TCP, 2TUPLE) | EFX_RX_HASH(IPV4_UDP, 2TUPLE) |
EFX_RX_HASH(IPV4, 2TUPLE).
At the same time, on ultra-low-latency firmware, the common code will
never report support for any UDP tuple to the client driver. That is,
in the same example, the driver will use EFX_RX_HASH(IPV4_TCP, 2TUPLE) |
EFX_RX_HASH(IPV4, 2TUPLE). This input will not be recognised by
EF10-specific handler, and RSS_CONTEXT_SET_FLAGS_IN_TOEPLITZ_IPV4_EN
bit will not be set in the MCDI request.
In order to solve the problem, the patch removes conversion code
from chip-specific handlers and adds appropriate code to convert
EFX_RX_HASH() flags to their legacy counterparts to the common scale
mode set function. If the firmware does not support additional modes,
the function will convert generalised flags to legacy flags correctly
without any demand for UDP flags and pass the result to a chip-specific
handler.
Submitted by: Ivan Malov <ivan.malov at oktetlabs.ru>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18269
Andrew Rybchenko [Fri, 30 Nov 2018 07:07:20 +0000 (07:07 +0000)]
sfxge(4): simplify the code to parse RSS hash type
RSS mode bits can be accessed a lot easier in the hash
type value provided that the variable type is uint32_t.
The macro helper can be removed to enhance readability.
Submitted by: Ivan Malov <ivan.malov at oktetlabs.ru>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18268
Andrew Rybchenko [Fri, 30 Nov 2018 07:07:09 +0000 (07:07 +0000)]
sfxge(4): check buffer size for hash flags
The efx_rx_scale_hash_flags_get interface is unsafe, as it does not
have an argument for the size of the output buffer used to return
the flags. While the only caller currently supplies a sufficiently
large buffer, this should be checked at runtime to avoid writing
past the end of the buffer.
Submitted by: Ivan Malov <ivan.malov at oktetlabs.ru>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18267
Andrew Rybchenko [Fri, 30 Nov 2018 07:06:58 +0000 (07:06 +0000)]
sfxge(4): use simpler code to check hash algorithm type
The API which is used to list supported hash flags verifies
hash algorithm choice before writing the output. This check
is based on a switch() statement which has only two options
and no distinctive actions to be conducted for each of them.
Use simpler code instead of switch() to improve readability.
Submitted by: Ivan Malov <ivan.malov at oktetlabs.ru>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18266
Andrew Rybchenko [Fri, 30 Nov 2018 07:06:01 +0000 (07:06 +0000)]
sfxge(4): add API to inform libefx of hardware removal
The efx_nic_hw_unavailable() checks ensure that if the NIC hardware
has failed or has been physically removed then libefx will stop
further attempts to access the hardware.
Add an interface for libefx clients to force unavailability, so the
hardware is treated as dead or removed even if still physically present.
Submitted by: Andy Moreton <amoreton at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18261
Andrew Rybchenko [Fri, 30 Nov 2018 07:05:49 +0000 (07:05 +0000)]
sfxge(4): add routine to check for hardware presence
Add efx_nic_hw_unavailable() routine to check for hardware presence
before continuing with NIC operations.
Submitted by: Andy Moreton <amoreton at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18260
Andrew Rybchenko [Fri, 30 Nov 2018 07:05:23 +0000 (07:05 +0000)]
sfxge(4): add information if TSO workaround is required
In SF bug 61297 it's been confirmed that the hardware does not always
calculate the TCP checksum correctly with TSO sends.
The value of the Total Length field (IPv4) or Payload Length field
(IPv6) is the critical factor. We're sufficiently confident that if
these fields are zero then the checksum will be calculated correctly.
The information may be used by the drivers to check if the workaround is
required when FATSOv2 is implemented.
Submitted by: Mark Spender <mspender at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18258
Andrew Rybchenko [Fri, 30 Nov 2018 07:05:00 +0000 (07:05 +0000)]
sfxge(4): generalise EF10 NVRAM buffer interface
The SFN driver's PartitionControl WMI object requires an API to parse
and filter partition data in TLV format, particularly for the Dynamic
Config partition. The ef10_nvram_buffer functions provide this
functionality but are tied to use with license partition only.
Modify functions so they are applicable to all TLV partitions and add
functions to support in-place tag modification.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18256
Andrew Rybchenko [Fri, 30 Nov 2018 07:04:48 +0000 (07:04 +0000)]
sfxge(4): add accessor for default port mode
Extend efx_mcdi_get_port_modes() to optionally pass on the default
port mode field. This provides a more direct way of handling the case
where the dynamic config does not specify the port mode than the
alternative of a lookup table indexed by MCFW subtype.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18255
Andrew Rybchenko [Fri, 30 Nov 2018 07:04:37 +0000 (07:04 +0000)]
sfxge(4): add buffer editing functions to boot config
Functions to process the DHCP option list format used by the expansion
ROM config buffers, to support extracting and updating of individual
options.
The initial use case is the driver presenting the global and per-PF
options as separate items, with the driver implementing the
synchronization of global options across the configuration buffers
for all PFs.
Submitted by: Richard Houldsworth <rhouldsworth at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18254
Andrew Rybchenko [Fri, 30 Nov 2018 07:04:25 +0000 (07:04 +0000)]
sfxge(4): add API to retrieve sensor limits
Submitted by: Martin Harvey <mharvey at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18253
Andrew Rybchenko [Fri, 30 Nov 2018 05:54:30 +0000 (05:54 +0000)]
sfxge(4): add generated description of sensors
Description of sensors is generated from firmware sources.
Submitted by: Martin Harvey <mharvey at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18251
Andrew Rybchenko [Fri, 30 Nov 2018 05:54:19 +0000 (05:54 +0000)]
sfxge(4): remove probes when a Tx queue is too full
No need for probe messages when a TxQ is too full for a post to be done.
Existing drivers check if there is room in the queue before posting
descriptors, even though efx_tx_qdesc_post() does the check itself.
The new SFN Windows driver doesn't perform the check before calling
efx_tx_qdesc_post(), but that means these probes can get frequently
printed out. It's normal driver behaviour so there's no need to print
an error.
Submitted by: Mark Spender <mspender at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18250
Andrew Rybchenko [Fri, 30 Nov 2018 05:54:07 +0000 (05:54 +0000)]
sfxge(4): refactor monitors support
Remove obsolete monitor types since Falcon SFN4000 series adapters
no longer supported by libefx.
Rename MCDI monitors to be consistent with YML.
The code may be simplified and generalized since only MCDI monitors
remain.
Submitted by: Martin Harvey <mharvey at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18249
Andrew Rybchenko [Fri, 30 Nov 2018 05:50:01 +0000 (05:50 +0000)]
sfxge(4): move empty efsys definitions to EFX headers
Move empty definitions for platform-specific annotations from efsys.h
to EFX headers.
Submitted by: Martin Harvey <mharvey at solarflare.com>
Submitted by: Andrew Lee <alee at solarflare.com>
Sponsored by: Solarflare Communications, Inc.
Differential Revision: https://reviews.freebsd.org/D18248
Eric van Gyzen [Fri, 30 Nov 2018 04:59:43 +0000 (04:59 +0000)]
Include path for tmpfs objects in vm.objects sysctl
This applies the fix in r283924 to the vm.objects sysctl
added by r283624 so the output will include the vnode
information (i.e. path) for tmpfs objects.
Cy Schubert [Fri, 30 Nov 2018 04:15:42 +0000 (04:15 +0000)]
Clean up a redundant non-redefinition of IFNAMSIZ. IFNAMSIZ
is defined in net/if.h, therefore the condition is never met and
confusing to those who follow.
Eric van Gyzen [Fri, 30 Nov 2018 03:01:32 +0000 (03:01 +0000)]
_thr_setthreaded() cannot fail; change return type to void
Also remove logic to avoid unnecessary stores to the global variable.
Thread creation and destruction are heavy enough that any supposed savings
is in the noise.
Mateusz Guzik [Fri, 30 Nov 2018 00:45:10 +0000 (00:45 +0000)]
amd64: tidy up copying backwards in memmove
For non-ERMS case the code used handle possible trailing bytes with
movsb first and then followed it up with movsq. This also happened
to alter how calculations were done for other cases.
Handle the tail with regular movs, just like when copying forward.
Use leaq to calculate the right offset from the get go, instead of
doing separate add and sub.
This adjusts the offset for non-rep cases so that they can be used
to handle the tail.
Brooks Davis [Thu, 29 Nov 2018 21:00:56 +0000 (21:00 +0000)]
Add helper functions to copy strings into struct image_args.
Given a zeroed struct image_args with an allocated buf member,
exec_args_add_fname() must be called to install a file name (or NULL).
Then zero or more calls to exec_args_add_env() followed by zero or
more calls to exec_args_add_env(). exec_args_adjust_args() may be
called after args and/or env to allow an interpreter to be prepended to
the argument list.
To allow code reuse when adding arg and env variables, begin_envv
should be accessed with the accessor exec_args_get_begin_envv()
which handles the case when no environment entries have been added.
Use these functions to simplify exec_copyin_args() and
freebsd32_exec_copyin_args().
Maxim Sobolev [Thu, 29 Nov 2018 19:28:01 +0000 (19:28 +0000)]
Replace hand-crafted naive byte-by-byte zero block detection routine
with macro based around memcmp(). The latter is expected to be some
8 times faster on a modern 64-bit architectures.
In practice, throughput of doing conv=sparse from /dev/zero to /dev/null
went up some 5-fold here from 1.9GB/sec to 9.7GB/sec with this change
(bs=128k).
If BENEATH is specified, always latch the topping directory vnode.
It is possible that we started with a relative path but during the
lookup, found an absolute symlink. In this case, BENEATH handling
code needs the latch, but it is too late to calculate it.
While there, somewhat improve the assertions. Clear the NI_LCF_LATCH
flag when the latch vnode is released, so that asserts know the state.
Assert that there is a latch if we entered beneath+abs path mode,
after the starting point is processed.
Reported by: wulf
With more input from: pho
Sponsored by: The FreeBSD Foundation
Emmanuel Vadot [Thu, 29 Nov 2018 19:06:05 +0000 (19:06 +0000)]
extres: clk: Fix clk_set_assigned
ofw_bus_parse_xref_list_get_length doesn't returns the number of elements, fix this.
While here when setting the clock to the assigned freqeuncy, allow the clock
driver to round down or up the frequency as sometimes the exact frequency cannot
be obtain.
Mark Johnston [Thu, 29 Nov 2018 16:31:01 +0000 (16:31 +0000)]
Update the free page count when blacklisting pages.
Otherwise the free page count will not accurately reflect the physical
page allocator's state. On 11 this can trigger panics in
vm_page_alloc() since the allocator state and free page count are
updated atomically and we expect them to stay in sync. On 12 the
bug would manifest as threads looping in vm_page_alloc().
PR: 231296
Reported by: mav, wollman, Rainer Duffner, Josh Gitlin
Reviewed by: alc, kib, mav
MFC after: 3 days
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18374