Pavel Begunkov says:
====================
io_uring zerocopy send
The patchset implements io_uring zerocopy send. It works with both registered
and normal buffers, mixing is allowed but not recommended. Apart from usual
request completions, just as with MSG_ZEROCOPY, io_uring separately notifies
the userspace when buffers are freed and can be reused (see API design below),
which is delivered into io_uring's Completion Queue. Those "buffer-free"
notifications are not necessarily per request, but the userspace has control
over it and should explicitly attaching a number of requests to a single
notification. The series also adds some internal optimisations when used with
registered buffers like removing page referencing.
From the kernel networking perspective there are two main changes. The first
one is passing ubuf_info into the network layer from io_uring (inside of an
in kernel struct msghdr). This allows extra optimisations, e.g. ubuf_info
caching on the io_uring side, but also helps to avoid cross-referencing
and synchronisation problems. The second part is an optional optimisation
removing page referencing for requests with registered buffers.
Benchmarking UDP with an optimised version of the selftest (see [1]), which
sends a bunch of requests, waits for completions and repeats. "+ flush" column
posts one additional "buffer-free" notification per request, and just "zc"
doesn't post buffer notifications at all.
NIC (requests / second):
IO size | non-zc | zc | zc + flush
4000 | 495134 | 606420 (+22%) | 558971 (+12%)
1500 | 551808 | 577116 (+4.5%) | 565803 (+2.5%)
1000 | 584677 | 592088 (+1.2%) | 560885 (-4%)
600 | 596292 | 598550 (+0.4%) | 555366 (-6.7%)
dummy (requests / second):
IO size | non-zc | zc | zc + flush
8000 | 1299916 | 2396600 (+84%) | 2224219 (+71%)
4000 | 1869230 | 2344146 (+25%) | 2170069 (+16%)
1200 | 2071617 | 2361960 (+14%) | 2203052 (+6%)
600 | 2106794 | 2381527 (+13%) | 2195295 (+4%)
Previously it also brought a massive performance speedup compared to the
msg_zerocopy tool (see [3]), which is probably not super interesting. There
is also an additional bunch of refcounting optimisations that was omitted from
the series for simplicity and as they don't change the picture drastically,
they will be sent as follow up, as well as flushing optimisations closing the
performance gap b/w two last columns.
For TCP on localhost (with hacks enabling localhost zerocopy) and including
additional overhead for receive:
IO size | non-zc | zc
1200 | 4174 | 4148
4096 | 7597 | 11228
Using a real NIC 1200 bytes, zc is worse than non-zc ~5-10%, maybe the
omitted optimisations will somewhat help, should look better for 4000,
but couldn't test properly because of setup problems.
Links:
liburing (benchmark + tests):
[1] https://github.com/isilence/liburing/tree/zc_v4
kernel repo:
[2] https://github.com/isilence/linux/tree/zc_v4
RFC v1:
[3] https://lore.kernel.org/io-uring/cover.1638282789.git.asml.silence@gmail.com/
RFC v2:
https://lore.kernel.org/io-uring/cover.1640029579.git.asml.silence@gmail.com/
Net patches based:
git@github.com:isilence/linux.git zc_v4-net-base
or
https://github.com/isilence/linux/tree/zc_v4-net-base
API design overview:
The series introduces an io_uring concept of notifactors. From the userspace
perspective it's an entity to which it can bind one or more requests and then
requesting to flush it. Flushing a notifier makes it impossible to attach new
requests to it, and instructs the notifier to post a completion once all
requests attached to it are completed and the kernel doesn't need the buffers
anymore.
Notifications are stored in notification slots, which should be registered as
an array in io_uring. Each slot stores only one notifier at any particular
moment. Flushing removes it from the slot and the slot automatically replaces
it with a new notifier. All operations with notifiers are done by specifying
an index of a slot it's currently in.
When registering a notification the userspace specifies a u64 tag for each
slot, which will be copied in notification completion entries as
cqe::user_data. cqe::res is 0 and cqe::flags is equal to wrap around u32
sequence number counting notifiers of a slot.
====================
Link: https://lore.kernel.org/r/cover.1657643355.git.asml.silence@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Teach tcp how to use external ubuf_info provided in msghdr and
also prepare it for managed frags by sprinkling
skb_zcopy_downgrade_managed() when it could mix managed and not managed
frags.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Teach ipv6/udp how to use external ubuf_info provided in msghdr and
also prepare it for managed frags by sprinkling
skb_zcopy_downgrade_managed() when it could mix managed and not managed
frags.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Teach ipv4/udp how to use external ubuf_info provided in msghdr and
also prepare it for managed frags by sprinkling
skb_zcopy_downgrade_managed() when it could mix managed and not managed
frags.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Managed pages contain pinned userspace pages and controlled by upper
layers, there is no need in tracking skb->pfmemalloc for them. Introduce
a helper for filling frags but ignoring page tracking, it'll be needed
later.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some users like io_uring can do page pinning more efficiently, so we
want a way to delegate referencing to other subsystems. For that add
a new flag called SKBFL_MANAGED_FRAG_REFS. When set, skb doesn't hold
page references and upper layers are responsivle to managing page
lifetime.
It's allowed to convert skbs from managed to normal by calling
skb_zcopy_downgrade_managed(). The function will take all needed
page references and clear the flag. It's needed, for instance,
to avoid mixing managed modes.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add support for custom iov_iter handling to msghdr. The idea is that
in-kernel subsystems want control over how an SG is split.
Signed-off-by: David Ahern <dsahern@kernel.org>
[pavel: move callback into msghdr]
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Make possible for network in-kernel callers like io_uring to pass in a
custom ubuf_info by setting it in a new field of struct msghdr.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Tony Nguyen says:
====================
100GbE Intel Wired LAN Driver Updates 2022-07-15
This series contains updates to ice driver only.
Ani updates feature restriction for devices that don't support external
time stamping.
Zhuo Chen removes unnecessary call to pci_aer_clear_nonfatal_status().
* '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue:
ice: Remove pci_aer_clear_nonfatal_status() call
ice: Add EXTTS feature to the feature bitmap
====================
Link: https://lore.kernel.org/r/20220715214642.2968799-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The port fields in the ethool flow structures are defined
to be __be16 types, so sparse is showing issues where these
are being passed to htons(). Fix these warnings by passing
them to be16_to_cpu() instead.
These are being used in netdev_dbg() so should only effect
anyone doing debug.
Fixes the following sparse warnings:
drivers/net/ethernet/cadence/macb_main.c:3366:9: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3366:9: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3366:9: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3419:25: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3419:25: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3419:25: warning: cast from restricted __be16
drivers/net/ethernet/cadence/macb_main.c:3419:25: warning: cast from restricted __be16
Signed-off-by: Ben Dooks <ben.dooks@sifive.com>
Link: https://lore.kernel.org/r/20220715173009.526126-1-ben.dooks@sifive.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Jiri Pirko says:
====================
devlink: prepare mlxsw and netdevsim for locked reload
This is preparation patchset to be able to eventually make a switch and
make reload cmd to take devlink->lock as the other commands do.
This patchset is preparing 2 major users of devlink API - mlxsw and
netdevsim. The sets of functions are similar, therefore taking care of
both here.
====================
Link: https://lore.kernel.org/r/20220716110241.3390528-1-jiri@resnulli.us
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove locked versions of functions that are no longer used by anyone.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare for devlink reload being called with devlink->lock held and
convert the netdevsim driver to use unlocked devlink API during init and
fini flows. Take devl_lock() in reload_down() and reload_up() ops in the
meantime before reload cmd is converted to take the lock itself.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add unlocked variants of devlink_region_create/destroy() functions
to be used in drivers called-in with devlink->lock held.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare for devlink reload being called with devlink->lock held and
convert the mlxsw driver to use unlocked devlink API during init and
fini flows. Take devl_lock() in reload_down() and reload_up() ops in the
meantime before reload cmd is converted to take the lock itself.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Tested-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add unlocked variants of devlink_dpipe*() functions to be used
in drivers called-in with devlink->lock held.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add unlocked variants of devlink_sb*() functions to be used
in drivers called-in with devlink->lock held.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add unlocked variants of devlink_resource*() functions to be used
in drivers called-in with devlink->lock held.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add unlocked variants of devl_trap*() functions to be used in drivers
called-in with devlink->lock held.
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We don't want to list every single ubuf_info callback in
skb_orphan_frags(), add a flag controlling the behaviour.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We should not append MSG_ZEROCOPY requests to skbuff with non
MSG_ZEROCOPY ubuf_info, they might be not compatible.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Even when zerocopy transmission is requested and possible,
__ip_append_data() will still copy a small chunk of data just because it
allocated some extra linear space (e.g. 128 bytes). It wastes CPU cycles
on copy and iter manipulations and also misalignes potentially aligned
data. Avoid such copies. And as a bonus we can allocate smaller skb.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Even when zerocopy transmission is requested and possible,
__ip_append_data() will still copy a small chunk of data just because it
allocated some extra linear space (e.g. 148 bytes). It wastes CPU cycles
on copy and iter manipulations and also misalignes potentially aligned
data. Avoid such copies. And as a bonus we can allocate smaller skb.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use netif_napi_add_tx() for NAPI in Tx direction instead of the regular
netif_napi_add() function.
Signed-off-by: Sieng-Piaw Liew <liew.s.piaw@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes the of_match_ptr() pointer when dereferencing the
ksz_dt_ids which produce the unused variable warning.
Reported-by: kernel test robot <lkp@intel.com>
Suggested-by: Arnd Bergmann <arnd@kernel.org>
Signed-off-by: Arun Ramadoss <arun.ramadoss@microchip.com>
Reviewed-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jakub Kicinski says:
====================
tls: rx: avoid skb_cow_data()
TLS calls skb_cow_data() on the skb it received from strparser
whenever it needs to hold onto the skb with the decrypted data.
(The alternative being decrypting directly to a user space buffer
in whic case the input skb doesn't get modified or used after.)
TLS needs the decrypted skb:
- almost always with TLS 1.3 (unless the new NoPad is enabled);
- when user space buffer is too small to fit the record;
- when BPF sockmap is enabled.
Most of the time the skb we get out of strparser is a clone of
a 64kB data unit coalsced by GRO. To make things worse skb_cow_data()
tries to output a linear skb and allocates it with GFP_ATOMIC.
This occasionally fails even under moderate memory pressure.
This patch set rejigs the TLS Rx so that we don't expect decryption
in place. The decryption handlers return an skb which may or may not
be the skb from strparser. For TLS 1.3 this results in a 20-30%
performance improvement without NoPad enabled.
v2: rebase after 3d8c51b25a ("net/tls: Check for errors in tls_device_init")
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
We currently CoW Rx skbs whenever we can't decrypt to a user
space buffer. The skbs can be enormous (64kB) and CoW does
a linear alloc which has a strong chance of failing under
memory pressure. Or even without, skb_cow_data() assumes
GFP_ATOMIC.
Allocate a new frag'd skb and decrypt into it. We finally
take advantage of the decrypted skb getting returned via
darg.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The "zero-copy" path in SW TLS will engage either for no skbs or
for all but last. If the recvmsg parameters are right and the
socket can do ZC we'll ZC until the iterator can't fit a full
record at which point we'll decrypt one more record and copy
over the necessary bits to fill up the request.
The only reason we hold onto the ZC skbs which went thru the async
path until the end of recvmsg() is to count bytes. We need an accurate
count of zc'ed bytes so that we can calculate how much of the non-zc'd
data to copy. To allow freeing input skbs on the ZC path count only
how much of the list we'll need to consume.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Async crypto currently benefits from the fact that we decrypt
in place. When we allow input and output to be different skbs
we will have to hang onto the input while we move to the next
record. Clone the inputs and keep them on a list.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Async crypto TLS Rx currently waits for crypto to be done
in order to strip the TLS header and tailer. Simplify
the code by moving the pointers immediately, since only
TLS 1.2 is supported here there is no message padding.
This simplifies the decryption into a new skb in the next
patch as we don't have to worry about input vs output
skb in the decrypt_done() handler any more.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of using ctx->recv_pkt after decryption read the skb
from darg.skb. This moves the decision of what the "output skb"
is to the decrypt handlers. For now after decrypt handler returns
successfully ctx->recv_pkt is simply moved to darg.skb, but it
will change soon.
Note that tls_decrypt_sg() cannot clear the ctx->recv_pkt
because it gets called to re-encrypt (i.e. by the device offload).
So we need an awkward temporary if() in tls_rx_one_record().
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Callers always pass ctx->recv_pkt into decrypt_skb_update(),
and it propagates it to its callees. This may give someone
the false impression that those functions can accept any valid
skb containing a TLS record. That's not the case, the record
sequence number is read from the context, and they can only
take the next record coming out of the strp.
Let the functions get the skb from the context instead of
passing it in. This will also make it cleaner to return
a different skb than ctx->recv_pkt as the decrypted one
later on.
Since we're touching the definition of decrypt_skb_update()
use this as an opportunity to rename it.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
I already forgot to transform darg from input to output
semantics once on the NIC inline crypto fastpath. To
avoid this happening again create a device equivalent
of decrypt_internal(). A function responsible for decryption
and transforming darg.
While at it rename decrypt_internal() to a hopefully slightly
more meaningful name.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
We no longer allow a decrypted skb to remain linked to ctx->recv_pkt.
Anything on the list is decrypted, anything on ctx->recv_pkt needs
to be decrypted.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Detach the skb from ctx->recv_pkt after decryption is done,
even if we can't consume it.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
I thought that having the skb either always on the ctx->rx_list
or ctx->recv_pkt will simplify the handling, as we would not
have to remember to flip it from one to the other on exit paths.
This became a little harder to justify with the fix for BPF
sockmaps. Subsequent changes will make the situation even worse.
Queue the skbs only when really needed.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
recvmsg() in TLS gets data from the skb list (rx_list) or fresh
skbs we read from TCP via strparser. The former holds skbs which were
already decrypted for peek or decrypted and partially consumed.
tls_wait_data() only notices appearance of fresh skbs coming out
of TCP (or psock). It is possible, if there is a concurrent call
to peek() and recv() that the peek() will move the data from input
to rx_list without recv() noticing. recv() will then read data out
of order or never wake up.
This is not a practical use case/concern, but it makes the self
tests less reliable. This patch solves the problem by allowing
only one reader in.
Because having multiple processes calling read()/peek() is not
normal avoid adding a lock and try to fast-path the single reader
case.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Wen Gu says:
====================
net/smc: Introduce virtually contiguous buffers for SMC-R
On long-running enterprise production servers, high-order contiguous
memory pages are usually very rare and in most cases we can only get
fragmented pages.
When replacing TCP with SMC-R in such production scenarios, attempting
to allocate high-order physically contiguous sndbufs and RMBs may result
in frequent memory compaction, which will cause unexpected hung issue
and further stability risks.
So this patch set is aimed to allow SMC-R link group to use virtually
contiguous sndbufs and RMBs to avoid potential issues mentioned above.
Whether to use physically or virtually contiguous buffers can be set
by sysctl smcr_buf_type.
Note that using virtually contiguous buffers will bring an acceptable
performance regression, which can be mainly divided into two parts:
1) regression in data path, which is brought by additional address
translation of sndbuf by RNIC in Tx. But in general, translating
address through MTT is fast. According to qperf test, this part
regression is basically less than 10% in latency and bandwidth.
(see patch 5/6 for details)
2) regression in buffer initialization and destruction path, which is
brought by additional MR operations of sndbufs. But thanks to link
group buffer reuse mechanism, the impact of this kind of regression
decreases as times of buffer reuse increases.
Patch set overview:
- Patch 1/6 and 2/6 mainly about simplifying and optimizing DMA sync
operation, which will reduce overhead on the data path, especially
when using virtually contiguous buffers;
- Patch 3/6 and 4/6 introduce a sysctl smcr_buf_type to set the type
of buffers in new created link group;
- Patch 5/6 allows SMC-R to use virtually contiguous sndbufs and RMBs,
including buffer creation, destruction, MR operation and access;
- patch 6/6 extends netlink attribute for buffer type of SMC-R link group;
v1->v2:
- Patch 5/6 fixes build issue on 32bit;
- Patch 3/6 adds description of new sysctl in smc-sysctl.rst;
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Extend SMC-R link group netlink attribute SMC_GEN_LGR_SMCR.
Introduce SMC_NLA_LGR_R_BUF_TYPE to show the buffer type of
SMC-R link group.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On long-running enterprise production servers, high-order contiguous
memory pages are usually very rare and in most cases we can only get
fragmented pages.
When replacing TCP with SMC-R in such production scenarios, attempting
to allocate high-order physically contiguous sndbufs and RMBs may result
in frequent memory compaction, which will cause unexpected hung issue
and further stability risks.
So this patch is aimed to allow SMC-R link group to use virtually
contiguous sndbufs and RMBs to avoid potential issues mentioned above.
Whether to use physically or virtually contiguous buffers can be set
by sysctl smcr_buf_type.
Note that using virtually contiguous buffers will bring an acceptable
performance regression, which can be mainly divided into two parts:
1) regression in data path, which is brought by additional address
translation of sndbuf by RNIC in Tx. But in general, translating
address through MTT is fast.
Taking 256KB sndbuf and RMB as an example, the comparisons in qperf
latency and bandwidth test with physically and virtually contiguous
buffers are as follows:
- client:
smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\
-t 5 -vu tcp_{bw|lat}
- server:
smc_run taskset -c <cpu> qperf
[latency]
msgsize tcp smcr smcr-use-virt-buf
1 11.17 us 7.56 us 7.51 us (-0.67%)
2 10.65 us 7.74 us 7.56 us (-2.31%)
4 11.11 us 7.52 us 7.59 us ( 0.84%)
8 10.83 us 7.55 us 7.51 us (-0.48%)
16 11.21 us 7.46 us 7.51 us ( 0.71%)
32 10.65 us 7.53 us 7.58 us ( 0.61%)
64 10.95 us 7.74 us 7.80 us ( 0.76%)
128 11.14 us 7.83 us 7.87 us ( 0.47%)
256 10.97 us 7.94 us 7.92 us (-0.28%)
512 11.23 us 7.94 us 8.20 us ( 3.25%)
1024 11.60 us 8.12 us 8.20 us ( 0.96%)
2048 14.04 us 8.30 us 8.51 us ( 2.49%)
4096 16.88 us 9.13 us 9.07 us (-0.64%)
8192 22.50 us 10.56 us 11.22 us ( 6.26%)
16384 28.99 us 12.88 us 13.83 us ( 7.37%)
32768 40.13 us 16.76 us 16.95 us ( 1.16%)
65536 68.70 us 24.68 us 24.85 us ( 0.68%)
[bandwidth]
msgsize tcp smcr smcr-use-virt-buf
1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%)
2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%)
4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%)
8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%)
16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%)
32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%)
64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%)
128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%)
256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%)
512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%)
1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%)
2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%)
4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%)
8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%)
16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%)
32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%)
65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%)
2) regression in buffer initialization and destruction path, which is
brought by additional MR operations of sndbufs. But thanks to link
group buffer reuse mechanism, the impact of this kind of regression
decreases as times of buffer reuse increases.
Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R
buffer-related function obtained by bpftrace are as follows:
Function Phys-bufs Virt-bufs
smcr_new_buf_create() 67154 ns 79164 ns
smc_ib_buf_map_sg() 525 ns 928 ns
smc_ib_get_memory_region() 162294 ns 161191 ns
smc_wr_reg_send() 9957 ns 9635 ns
smc_ib_put_memory_region() 203548 ns 198374 ns
smc_ib_buf_unmap_sg() 508 ns 1158 ns
------------
Test environment notes:
1. Above tests run on 2 VMs within the same Host.
2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to
the each VM respectively.
3. VMs' vCPUs are binded to different physical CPUs, and the binded
physical CPUs are isolated by `isolcpus=xxx` cmdline.
4. NICs' queue number are set to 1.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces a new SMC-R specific element buf_type
in struct smc_link_group, for recording the value of sysctl
smcr_buf_type when link group is created.
New created link group will create and reuse buffers of the
type specified by buf_type.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces the sysctl smcr_buf_type for setting
the type of SMC-R sndbufs and RMBs.
Valid values includes:
- SMCR_PHYS_CONT_BUFS, which means use physically contiguous
buffers for better performance and is the default value.
- SMCR_VIRT_CONT_BUFS, which means use virtually contiguous
buffers in case of physically contiguous memory is scarce.
- SMCR_MIXED_BUFS, which means first try to use physically
contiguous buffers. If not available, then use virtually
contiguous buffers.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some CPU, such as Xeon, can guarantee DMA cache coherency.
So it is no need to use dma sync APIs to flush cache on such CPUs.
In order to avoid calling dma sync APIs on the IO path, use the
dma_need_sync to check whether smc_buf_desc needs dma sync when
creating smc_buf_desc.
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
