There are several mechanisms by which a vm_page reference is held,
preventing the page from being freed back to the page allocator. In
particular, holding the page's object lock is sufficient to prevent the
page from being freed; holding the busy lock or a wiring is sufficent as
well. These references are protected by the page lock, which must
therefore be acquired for many per-page operations. This results in
false sharing since the page locks are external to the vm_page
structures themselves and each lock protects multiple structures.
Transition to using an atomically updated per-page reference counter.
The object's reference is counted using a flag bit in the counter. A
second flag bit is used to atomically block new references via
pmap_extract_and_hold() while removing managed mappings of a page.
Thus, the reference count of a page is guaranteed not to increase if the
page is unbusied, unmapped, and the object's write lock is held. As
a consequence of this, the page lock no longer protects a page's
identity; operations which move pages between objects are now
synchronized solely by the objects' locks.
The vm_page_wire() and vm_page_unwire() KPIs are changed. The former
requires that either the object lock or the busy lock is held. The
latter no longer has a return value and may free the page if it releases
the last reference to that page. vm_page_unwire_noq() behaves the same
as before; the caller is responsible for checking its return value and
freeing or enqueuing the page as appropriate. vm_page_wire_mapped() is
introduced for use in pmap_extract_and_hold(). It fails if the page is
concurrently being unmapped, typically triggering a fallback to the
fault handler. vm_page_wire() no longer requires the page lock and
vm_page_unwire() now internally acquires the page lock when releasing
the last wiring of a page (since the page lock still protects a page's
queue state). In particular, synchronization details are no longer
leaked into the caller.
The change excises the page lock from several frequently executed code
paths. In particular, vm_object_terminate() no longer bounces between
page locks as it releases an object's pages, and direct I/O and
sendfile(SF_NOCACHE) completions no longer require the page lock. In
these latter cases we now get linear scalability in the common scenario
where different threads are operating on different files.
__FreeBSD_version is bumped. The DRM ports have been updated to
accomodate the KPI changes.
Reviewed by: jeff (earlier version)
Tested by: gallatin (earlier version), pho
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20486
Remove now-redundant items from toepcb and synq_entry and the code to
support them.
Let the driver calculate tx_align, rx_coalesce, and sndbuf by default.
Reviewed by: jhb@
MFC after: 1 week
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D21387
The hold_count and wire_count fields of struct vm_page are separate
reference counters with similar semantics. The remaining essential
differences are that holds are not counted as a reference with respect
to LRU, and holds have an implicit free-on-last unhold semantic whereas
vm_page_unwire() callers must explicitly determine whether to free the
page once the last reference to the page is released.
This change removes the KPIs which directly manipulate hold_count.
Functions such as vm_fault_quick_hold_pages() now return wired pages
instead. Since r328977 the overhead of maintaining LRU for wired pages
is lower, and in many cases vm_fault_quick_hold_pages() callers would
swap holds for wirings on the returned pages anyway, so with this change
we remove a number of page lock acquisitions.
No functional change is intended. __FreeBSD_version is bumped.
Reviewed by: alc, kib
Discussed with: jeff
Discussed with: jhb, np (cxgbe)
Tested by: pho (previous version)
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D19247
receive sockbuf's high water mark.
Calculate rx credits on the spot instead of tracking sbused/sb_cc and
rx_credits in the toepcb. The previous method worked when the high
water mark changed due to SB_AUTOSIZE but not when it was adjusted
directly (for example, by the soreserve in nfsrvd_addsock).
This fixes a connection hang while running iozone over an NFS mounted
share where nfsd's TCP sockets are being handled by t4_tom.
MFC after: 3 days
Sponsored by: Chelsio Communications
The declaration in tcp_var.h is still around so t4_tom continued to
compile but wouldn't load. A separate commit will fix tcp_var.h
Reported By: Dustin Marquess (dmarquess at gmail)
Sponsored by: Chelsio Communications
intended recipient of a CPL when it can't be determined solely from the
opcode. Retire the per-queue handlers for such CPLs in favor of the new
scheme.
Sponsored by: Chelsio Communications
Requests to modify the state of TLS connections need to be sent on the
same queue as TLS record transmit requests to ensure ordering.
However, in order to use the offload transmit queue in t4_set_tcb_field(),
the function needs to be updated to do proper flow control / credit
management when queueing a request to an offload queue. This required
passing a pointer to the toepcb itself to this function, so while here
remove the 'tid' and 'iqid' parameters and obtain those values from the
toepcb in t4_set_tcb_field() itself.
Submitted by: Harsh Jain @ Chelsio (original version)
Reviewed by: np
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D14871
- Change t4_ddp_mod_load() to return void instead of always returning
success. This avoids having to pretend to have proper support for
unloading when only part of t4_tom_mod_load() has run.
- If t4_register_uld() fails, don't invoke t4_tom_mod_unload() directly.
The module handling code in the kernel invokes MOD_UNLOAD on a module
whose MOD_LOAD fails with an error already.
Reviewed by: np (part of a larger patch)
MFC after: 1 month
Sponsored by: Chelsio Communications
This consolidates all of the DDP state in one place. Also, the code has
now been fixed to ensure that DDP state is only accessed for DDP
connections. This should not be a functional change but makes it cleaner
and easier to add state for other TOE socket modes in the future.
MFC after: 1 month
Sponsored by: Chelsio Communications
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using misidentified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
r324539 gathered some vnet decls into netinet/tcp_var.h, so that they
are now redundant in dev/cxgbe/tom/{t4_cpl_io.c,t4_ddp.c}. This triggers
gcc -Wredundant-decls.
Reviewed by: np
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D12674
To optimize the case of ping-ponging between two buffers, the DDP code
caches the last two buffers used keeping the pages wired and page pods
stored in the NIC's RAM. If a new aio_read() request uses one of the
same buffers, then the work of holding pages, etc. can be avoided.
However, the starting virtual address of an aio buffer was not saved,
only the page count, length, and initial page offset. Thus, an
aio_read() request could match a different buffer in the address
space. (Earlier during development vm_fault_hold_quick_pages() was
always called and the vm_page_t values were compared, but that was
eventually removed without being adequately replaced.) Fix by storing
the starting virtual address and comparing that (along with other
fields) to determine if a buffer can be reused.
MFC after: 3 days
Sponsored by: Chelsio Communications
Active Open:
- Save the socket's vnet at the time of the active open (t4_connect) and
switch to it when processing the reply (do_act_open_rpl or
do_act_establish).
Passive Open:
- Save the listening socket's vnet in the driver's listen_ctx and switch
to it when processing incoming SYNs for the socket.
- Reject SYNs that arrive on an ifnet that's not in the same vnet as the
listening socket.
CLIP (Compressed Local IPv6) table:
- Add only those IPv6 addresses to the CLIP that are in a vnet
associated with one of the card's ifnets.
Misc:
- Set vnet from the toepcb when processing TCP state transitions.
- The kernel sets the vnet when calling the driver's output routine
so t4_push_frames runs in proper vnet context already. One exception
is when incoming credits trigger tx within the driver's ithread. Set
the vnet explicitly in do_fw4_ack for that case.
MFC after: 3 days
Sponsored by: Chelsio Communications
related to "shared" CPLs.
a) Combine t4_set_tcb_field and t4_set_tcb_field_rpl into a single
function. Allow callers to direct the response to any iq. Tidy up
set_ulp_mode_iscsi while there to use names from t4_tcb.h instead of
magic constants.
b) Remove all CPL handler tables from struct adapter. This reduces its
size by around 2KB. All handlers are now registered at MOD_LOAD instead
of attach or some kind of initialization/activation. The registration
functions do not need an adapter parameter any more.
c) Add per-iq handlers to deal with CPLs whose destination cannot be
determined solely from the opcode. There are 2 such CPLs in use right
now: SET_TCB_RPL and L2T_WRITE_RPL. The base driver continues to send
filter and L2T_WRITEs over the mgmtq and solicits the reply on fwq.
t4_tom (including the DDP code) now uses the port's ctrlq to send
L2T_WRITEs and SET_TCB_FIELDs and solicits the reply on an ofld_rxq.
fwq and ofld_rxq have different handlers that know what kind of tid to
expect in the reply. Update t4_write_l2e and callers to to support any
wrq/iq combination.
Approved by: re@ (kib@)
Sponsored by: Chelsio Communications
File and disk-backed I/O requests store counts of read/written disk
blocks in each AIO job so that they can be charged to the thread that
completes an AIO request via aio_return() or aio_waitcomplete(). This
change extends AIO jobs to store counts of received/sent messages and
updates socket backends to set these counts accordingly. Note that
the socket backends are careful to only charge a single messages for
each AIO request even though a single request on a blocking socket might
invoke sosend or soreceive multiple times. This is to mimic the
resource accounting of synchronous read/write.
Adjust the UNIX socketpair AIO test to verify that the message resource
usage counts update accordingly for aio_read and aio_write.
Approved by: re (hrs)
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D6911
This reduces the size of kaiocb slightly. I've also added some generic
fields that other backends can use in place of the BIO-specific fields.
Change the socket and Chelsio DDP backends to use 'backend3' instead of
abusing _aiocb_private.status directly. This confines the use of
_aiocb_private to the AIO internals in vfs_aio.c.
Reviewed by: kib (earlier version)
Approved by: re (gjb)
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D6547
Chelsio's TCP offload engine supports direct DMA of received TCP payload
into wired user buffers. This feature is known as Direct-Data Placement.
However, to scale well the adapter needs to prepare buffers for DDP
before data arrives. aio_read() is more amenable to this requirement than
read() as applications often call read() only after data is available in
the socket buffer.
When DDP is enabled, TOE sockets use the recently added pru_aio_queue
protocol hook to claim aio_read(2) requests instead of letting them use
the default AIO socket logic. The DDP feature supports scheduling DMA
to two buffers at a time so that the second buffer is ready for use
after the first buffer is filled. The aio/DDP code optimizes the case
of an application ping-ponging between two buffers (similar to the
zero-copy bpf(4) code) by keeping the two most recently used AIO buffers
wired. If a buffer is reused, the aio/DDP code is able to reuse the
vm_page_t array as well as page pod mappings (a kind of MMU mapping the
Chelsio NIC uses to describe user buffers). The generation of the
vmspace of the calling process is used in conjunction with the user
buffer's address and length to determine if a user buffer matches a
previously used buffer. If an application queues a buffer for AIO that
does not match a previously used buffer then the least recently used
buffer is unwired before the new buffer is wired. This ensures that no
more than two user buffers per socket are ever wired.
Note that this feature is best suited to applications sending a steady
stream of data vs short bursts of traffic.
Discussed with: np
Relnotes: yes
Sponsored by: Chelsio Communications
handle_ddp_close() function in t4_ddp.c as the logic is similar
to handle_ddp_data(). This allows all knowledge of the special
DDP mbufs to be private to t4_ddp.c as well.
the no-longer existant sb_cc sockbuf member.
- Use sbavail() instead of sbused() in t4_soreceive_ddp() to match the
usage in soreceive_stream() on which it is based.
Discussed with: glebius (2)
sending not ready data:
o Add new flag to pru_send() flags - PRUS_NOTREADY.
o Add new protocol method pru_ready().
Sponsored by: Nginx, Inc.
Sponsored by: Netflix
o Introduce a notion of "not ready" mbufs in socket buffers. These
mbufs are now being populated by some I/O in background and are
referenced outside. This forces following implications:
- An mbuf which is "not ready" can't be taken out of the buffer.
- An mbuf that is behind a "not ready" in the queue neither.
- If sockbet buffer is flushed, then "not ready" mbufs shouln't be
freed.
o In struct sockbuf the sb_cc field is split into sb_ccc and sb_acc.
The sb_ccc stands for ""claimed character count", or "committed
character count". And the sb_acc is "available character count".
Consumers of socket buffer API shouldn't already access them directly,
but use sbused() and sbavail() respectively.
o Not ready mbufs are marked with M_NOTREADY, and ready but blocked ones
with M_BLOCKED.
o New field sb_fnrdy points to the first not ready mbuf, to avoid linear
search.
o New function sbready() is provided to activate certain amount of mbufs
in a socket buffer.
A special note on SCTP:
SCTP has its own sockbufs. Unfortunately, FreeBSD stack doesn't yet
allow protocol specific sockbufs. Thus, SCTP does some hacks to make
itself compatible with FreeBSD: it manages sockbufs on its own, but keeps
sb_cc updated to inform the stack of amount of data in them. The new
notion of "not ready" data isn't supported by SCTP. Instead, only a
mechanical substitute is done: s/sb_cc/sb_ccc/.
A proper solution would be to take away struct sockbuf from struct
socket and allow protocols to implement their own socket buffers, like
SCTP already does. This was discussed with rrs@.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
sb_cc member of struct sockbuf to a couple of inline functions:
sbavail() and sbused()
Right now they are equal, but once notion of "not ready socket buffer data",
will be checked in, they are going to be different.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
(I'm committing this on behalf of my colleagues in the Storage team
at Chelsio).
Submitted by: Sreenivasa Honnur <shonnur at chelsio dot com>
Sponsored by: Chelsio Communications.
the queue where to enqueue pages that are going to be unwired.
- Add stronger checks to the enqueue/dequeue for the pagequeues when
adding and removing pages to them.
Of course, for unmanaged pages the queue parameter of vm_page_unwire() will
be ignored, just as the active parameter today.
This makes adding new pagequeues quicker.
This change effectively modifies the KPI. __FreeBSD_version will be,
however, bumped just when the full cache of free pages will be
evicted.
Sponsored by: EMC / Isilon storage division
Reviewed by: alc
Tested by: pho
includes support for the NIC and TOE features of the 40G, 10G, and
1G/100M cards based on the T5.
The ASIC is mostly backward compatible with the Terminator 4 so cxgbe(4)
has been updated instead of writing a brand new driver. T5 cards will
show up as cxl (short for cxlgb) ports attached to the t5nex bus driver.
Sponsored by: Chelsio
Basically, this is automatic rx zero copy when feasible. TCP payload is
DMA'd directly into the userspace buffer described by the uio submitted
in soreceive by an application.
- Works with sockets that are being handled by the TCP offload engine
of a T4 chip (you need t4_tom.ko module loaded after cxgbe, and an
"ifconfig +toe" on the cxgbe interface).
- Does not require any modification to the application.
- Not enabled by default. Use hw.t4nex.<X>.toe.ddp="1" to enable it.
