This patch is the driver for NTB hardware in AMD SoCs (ported from Linux)
and enables the NTB infrastructure like Doorbells, Scratchpads and Memory
window in AMD SoC. This driver has been validated using ntb_transport and
if_ntb driver already available in FreeBSD.
Submitted by: Rajesh Kumar <rajesh1.kumar@amd.com>
MFC after: 1 month
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D18774
This allows replacing "sys/eventfilter.h" includes with "sys/_eventfilter.h"
in other header files (e.g., sys/{bus,conf,cpu}.h) and reduces header
pollution substantially.
EVENTHANDLER_DECLARE and EVENTHANDLER_LIST_DECLAREs were moved out of .c
files into appropriate headers (e.g., sys/proc.h, powernv/opal.h).
As a side effect of reduced header pollution, many .c files and headers no
longer contain needed definitions. The remainder of the patch addresses
adding appropriate includes to fix those files.
LOCK_DEBUG and LOCK_FILE_LINE_ARG are moved to sys/_lock.h, as required by
sys/mutex.h since r326106 (but silently protected by header pollution prior
to this change).
No functional change (intended). Of course, any out of tree modules that
relied on header pollution for sys/eventhandler.h, sys/lock.h, or
sys/mutex.h inclusion need to be fixed. __FreeBSD_version has been bumped.
Its a hack, we can't know/list all DMA engines, but this covers all
I/OAT of Xeon E5/E7 at least from Sandy Bridge till Skylake I saw.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
Remove unused and easy to misuse PNP macro parameter
Inspired by r338025, just remove the element size parameter to the
MODULE_PNP_INFO macro entirely. The 'table' parameter is now required to
have correct pointer (or array) type. Since all invocations of the macro
already had this property and the emitted PNP data continues to include the
element size, there is no functional change.
Mostly done with the coccinelle 'spatch' tool:
$ cat modpnpsize0.cocci
@normaltables@
identifier b,c;
expression a,d,e;
declarer MODULE_PNP_INFO;
@@
MODULE_PNP_INFO(a,b,c,d,
-sizeof(d[0]),
e);
@singletons@
identifier b,c,d;
expression a;
declarer MODULE_PNP_INFO;
@@
MODULE_PNP_INFO(a,b,c,&d,
-sizeof(d),
1);
$ rg -l MODULE_PNP_INFO -- sys | \
xargs spatch --in-place --sp-file modpnpsize0.cocci
(Note that coccinelle invokes diff(1) via a PATH search and expects diff to
tolerate the -B flag, which BSD diff does not. So I had to link gdiff into
PATH as diff to use spatch.)
Tinderbox'd (-DMAKE_JUST_KERNELS).
Approved by: re (glen)
I was not aware Warner was making or planning to make forward progress in
this area and have since been informed of that.
It's easy to apply/reapply when churn dies down.
Inspired by r338025, just remove the element size parameter to the
MODULE_PNP_INFO macro entirely. The 'table' parameter is now required to
have correct pointer (or array) type. Since all invocations of the macro
already had this property and the emitted PNP data continues to include the
element size, there is no functional change.
Mostly done with the coccinelle 'spatch' tool:
$ cat modpnpsize0.cocci
@normaltables@
identifier b,c;
expression a,d,e;
declarer MODULE_PNP_INFO;
@@
MODULE_PNP_INFO(a,b,c,d,
-sizeof(d[0]),
e);
@singletons@
identifier b,c,d;
expression a;
declarer MODULE_PNP_INFO;
@@
MODULE_PNP_INFO(a,b,c,&d,
-sizeof(d),
1);
$ rg -l MODULE_PNP_INFO -- sys | \
xargs spatch --in-place --sp-file modpnpsize0.cocci
(Note that coccinelle invokes diff(1) via a PATH search and expects diff to
tolerate the -B flag, which BSD diff does not. So I had to link gdiff into
PATH as diff to use spatch.)
Tinderbox'd (-DMAKE_JUST_KERNELS).
We may not have enough contiguous memory later, when NTB connection get
established. It is quite likely that NTB windows are symmetric and this
allocation remain, but even if not, we will just reallocate it later.
MFC after: 2 weeks
I suppose it should make this code NUMA-aware with recent NUMA drop-in,
trying to allocate shared memory buffers from domain closer to NT-bridge.
MFC after: 2 weeks
When enabled by EEPROM, use it to relax translation address/size alignment
requirements for BAR2 window by 128 or 256 times.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
An eventual devd(8) or other component should be able to scan buses and
automatically load drivers that match device ids described in this metadata.
Reviewed by: imp
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D12364
There is no big need to burn CPU if other side may be not there yet. For
example, the PLX hardware by default enables the NTB link up on reset, not
dependig on driver to do it. In case of Intel hardware this also reduces
race between MSI-X workaround negotiation and upper layers, using the same
scratchpad registers in different time.
MFC after: 12 days
In theory this allows to avoid one more expensive doorbell register read
later in some scenarios. But in practice it also significantly increases
packet rate on PLX hardware, that I can't explain yet, possibly work-
arounding some interrupt delays.
MFC after: 13 days
Sponsored by: iXsystems, Inc.
It allows application driver get initial link state without racing with
hardware interrupts, thanks to the context rmlock held here.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
Those events may be reported as soon as callback is registered, if the link
is enabled by hardware or some other application.
While there, clean link_is_up variable on link down event.
MFC after: 1 week
This driver supports both NTB-to-NTB and NTB-to-Root Port modes (though
the second with predictable complications on hot-plug and reboot events).
I tested it with PEX 8717 and PEX 8733 chips, but expect it should work
with many other compatible ones too. It supports up to two NT bridges
per chip, each of which can have up to 2 64-bit or 4 32-bit memory windows,
6 or 12 scratchpad registers and 16 doorbells. There are also 4 DMA engines
in those chips, but they are not yet supported.
While there, rename Intel NTB driver from generic ntb_hw(4) to more specific
ntb_hw_intel(4), so now it is on par with this new ntb_hw_plx(4) driver and
alike to Linux naming.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
This fixes interrupt storms on hardware using legacy level-triggered
interrupts, since doorbell processing could take time after interrupt
handler completion, that triggered extra interrupts in a loop.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
Since the doorbell bit is already set when interrupt handler is called,
the event was not propagated to upper layer. It was working normally
because present code was not using masking actively, but that is going
to change.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
I believe it never worked correctly for more the one queue even in Linux.
This fixes case when one of consumer drivers is not loaded on one side,
but its queues still announced as ready if something else brought link up.
While there, remove some pointless NULL checks.
New design allows to attach multiple consumers to ntb_transport(4) instance.
Previous design obtained from Linux theoretically allowed that, but was not
practically usable (Linux also has only one consumer driver now).
New design allows hardware resources to be split between several consumers.
For example, one BAR can be dedicated for remote memory access, while other
resources can be used for packet transport for virtual Ethernet interface.
And even without resource split, this code allows to specify which consumer
driver should attach the hardware.
From some points this makes the code even closer to Linux one, even though
Linux does not provide the described flexibility.
Calling it earlier increases the window when MSIX info may change.
This change does not solve the problem completely, but seems logical.
Complete solution should probably include link reset in case of MSIX
remap to trigger new negotiation, but we have no way to get notified
about that now.
I don't know what errata is mentioned there, I was unable to find it, but
setting limit before the base simply does not work at all. According to
specification attempt to set limit out of the present window range resets
it to zero, effectively disabling it. And that is what I see in practice.
Fixing this properly disables access for remote side to our memory until
respective xlat is negotiated and set. As I see, Linux does the same.
At that point link is quite likely not established yet, so messing with
scratch registers is premature there. Original commit message mentioned
code diff reduction from Linux, but this line is not present in Linux now.
For some reason hack with sending MSI-X interrupts by writing to remote
LAPIC memory works only for 32-bit BARs, that are available only if split
BARs mode is enabled in BIOS. If it is not, complain loudly and fall back
to less efficient workaround.
For compatibility reasons make driver not report any checksum offload by
default, since there is indeed none. But if administrator knows that
interface is used only for local traffic, he can enable fake checksum
offload manually on both sides to save some CPU cycles, since the data
are already protected by CRC32 of PCIe link.
Sponsored by: iXsystems, Inc.
This allows at least first three doorbells to work very close to normal
hardware, properly signaling events to upper layers without spurious or
lost events. Doorbells above the first three may still report spurious
events due to lack of reliable information, but they are rarely used.
