Re: [RFC PATCH v3 00/11] virtio/vsock: experimental zerocopy receive
From: Stefano Garzarella
Date: Fri Nov 11 2022 - 08:48:34 EST
Hi Arseniy,
maybe we should start rebasing this series on the new support for
skbuff:
https://lore.kernel.org/lkml/20221110171723.24263-1-bobby.eshleman@xxxxxxxxxxxxx/
CCing Bobby to see if it's easy to integrate since you're both changing
the packet allocation.
On Sun, Nov 06, 2022 at 07:33:41PM +0000, Arseniy Krasnov wrote:
INTRODUCTION
Hello,
This is experimental patchset for virtio vsock zerocopy receive.
It was inspired by TCP zerocopy receive by Eric Dumazet. This API uses
same idea:call 'mmap()' on socket's descriptor,then call 'getsockopt()'
to fill provided vma area with pages of virtio receive buffers. After
received data was processed by user, pages must be freed by 'madvise()'
call with MADV_DONTNEED flag set(but if user will not call 'madvise()',
next 'getsockopt()' will fail).
DETAILS
Here is how mapping with mapped pages looks exactly: first page
contains information about mapped data buffers. At zero offset mapping
contains special data structure:
struct virtio_vsock_usr_hdr_pref {
u32 poll_value;
u32 hdr_num;
};
This structure contains two fields:
'poll_value' - shows that current socket has data to read.When socket's
intput queue is empty, 'poll_value' is set to 0 by kernel. When input
queue has some data, 'poll_value' is set to 1 by kernel. When socket is
closed for data receive, 'poll_value' is ~0.This tells user that "there
will be no more data,continue to call 'getsockopt()' until you'll find
'hdr_num' == 0".User spins on it in userspace, without calling 'poll()'
system call(of course, 'poll()' is still working).
'hdr_num' - shows number of mapped pages with data which starts from
second page of this mappined.
NOTE:
This version has two limitations:
1) One mapping per socket is supported. It is implemented by adding
'struct page*' pointer to 'struct virtio_vsock' structure (first
page of mapping, which contains 'virtio_vsock_usr_hdr_pref').But,
I think, support for multiple pages could be implemented by using
something like hash table of such pages, or more simple, just use
first page of mapping as headers page by default. Also I think,
number of such pages may be controlled by 'setsockop()'.
2) After 'mmap()' call,it is impossible to call 'mmap()' again, even
after calling 'madvise()'/'munmap()' on the whole mapping.This is
because socket can't handle 'munmap()' calls(as there is no such
callback in 'proto_ops'),thus polling page exists until socket is
opened.
After 'virtio_vsock_usr_hdr_pref' object, first page contains array of
trimmed virtio vsock packet headers (in contains only length of data on
the corresponding page and 'flags' field):
struct virtio_vsock_usr_hdr {
uint32_t length;
uint32_t flags;
};
Field 'length' allows user to know exact size of payload within each
sequence of pages and field 'flags' allows to process SOCK_SEQPACKET
flags(such as message bounds or record bounds).All other pages are data
pages from virtio queue.
Page 0 Page 1 Page N
[ pref hdr0 .. hdrN ][ data ] .. [ data ]
| | ^ ^
| | | |
| *-------|-----------*
| |
*----------------*
Of course, single header could represent array of pages (when
packet's buffer is bigger than one page).So here is example of detailed
mapping layout for some set of packages. Lets consider that we have the
following sequence of packages:56 bytes, 4096 bytes and 8200 bytes. All
pages: 0,1,2,3,4 and 5 will be inserted to user's vma.
Page 0: [[ pref ][ hdr0 ][ hdr 1 ][ hdr 2 ][ hdr 3 ] ... ]
Page 1: [ 56 ]
Page 2: [ 4096 ]
Page 3: [ 4096 ]
Page 4: [ 4096 ]
Page 5: [ 8 ]
Page 0 contains only array of headers:
'pref' is 'struct virtio_vsock_usr_hdr_pref'.
'hdr0' has 56 in length field.
'hdr1' has 4096 in length field.
'hdr2' has 8200 in length field.
'hdr3' has 0 in length field(this is end of data marker).
Page 1 corresponds to 'hdr0' and has only 56 bytes of data.
Page 2 corresponds to 'hdr1' and filled with data.
Page 3 corresponds to 'hdr2' and filled with data.
Page 4 corresponds to 'hdr2' and filled with data.
Page 5 corresponds to 'hdr2' and has only 8 bytes of data.
pref will be the following: poll_value = 1, hdr_num = 5
This patchset also changes packets allocation way: current uses
only 'kmalloc()' to create data buffer. Problem happens when we try to
map such buffers to user's vma - kernel restricts to map slab pages
to user's vma(as pages of "not large" 'kmalloc()' allocations have flag
PageSlab set and "not large" could be bigger than one page).So to avoid
this, data buffers now allocated using 'alloc_pages()' call.
DIFFERENCE WITH TCP
As this feature uses same approach as for TCP protocol,here are
some difference between both version(from user's POV):
1) For 'getsockopt()':
- This version passes only address of mapping.
- TCP passes special structure to 'getsockopt()'. In addition to the
address of mapping in contains 'length' and 'recv_skip_hint'.First
means size of data inside mapping(out param, set by kernel).Second
has bool type, if it is true, then user must dequeue rest of data
using 'read()' syscall(e.g. it is out parameter also).
2) Mapping structure:
- This version uses first page of mapping for meta data and rest of
pages for data.
- TCP version uses whole mapping for data only.
3) Data layout:
- This version inserts virtio buffers to mapping, so each buffer may
be filled partially. To get size of payload in every buffer, first
mapping's page must be used(see 2).
- TCP version inserts pages of single skb.
*Please, correct me if I made some mistake in TCP zerocopy description.
Thank you for the description. Do you think it would be possible to try
to do the same as TCP?
Especially now that we should support skbuff.
TESTS
This patchset updates 'vsock_test' utility: two tests for new
feature were added. First test covers invalid cases.Second checks valid
transmission case.
Thank you, I really appreciate you adding new tests each time! Great
job!
BENCHMARKING
For benchmakring I've created small test utility 'vsock_rx_perf'.
It works in client/server mode. When client connects to server, server
starts sending specified amount of data to client(size is set as input
argument). Client reads data and waits for next portion of it. In client
mode, dequeue logic works in three modes: copy, zerocopy and zerocopy
with user polling.
Cool, thanks for adding it in this series.
1) In copy mode client uses 'read()' system call.
2) In zerocopy mode client uses 'mmap()'/'getsockopt()' to dequeue data
and 'poll()' to wait data.
3) In zerocopy mode + user polling client uses 'mmap()'/'getsockopt()',
but to wait data it polls shared page(e.g. busyloop).
Here is usage:
-c <cid> Peer CID to connect to(if run in client mode).
-m <megabytes> Number of megabytes to send.
-b <bytes> Size of RX/TX buffer(or mapping) in pages.
-r <bytes> SO_RCVLOWAT value in bytes(if run in client mode).
-v <bytes> peer credit.
-s Run as server.
-z [n|y|u] Dequeue mode.
n - copy mode. 1) above.
y - zero copy mode. 2) above.
u - zero copy mode + user poll. 3) above.
Utility produces the following output:
1) In server mode it prints number of sec spent for whole tx loop.
2) In client mode it prints several values:
* Number of sec, spent for whole rx loop(including 'poll()').
* Number of sec, spend in dequeue system calls:
In case of '-z n' it will be time in 'read()'.
In case of '-z y|u' it will be time in 'getsockopt()' + 'madvise()'.
* Number of wake ups with POLLIN flag set(except '-z u' mode).
* Average time(ns) in single dequeue iteration(e.g. divide second
value by third).
Idea of test is to compare zerocopy approach and classic copy, as it is
clear, that to dequeue some "small" amount of data, copy must be better,
because syscall with 'memcpy()' for 1 byte(for example) is just nothing
against two system calls, where first must map at least one page, while
second will unmap it.
Test was performed with the following settings:
1) Total size of data to send is 2G(-m argument).
2) Peer's buffer size is changed to 2G(-v argument) - this is needed to
avoid stalls of sender to wait for enough credit.
3) Both buffer size(-b) and SO_RCVLOWAT(-r) are used to control number
of bytes to dequeue in single loop iteration. Buffer size limits max
number of bytes to read, while SO_RCVLOWAT won't allow user to get
too small number of bytes.
4) For sender, tx buffer(which is passed to 'write()') size is 16Mb. Of
course we can set it to peer's buffer size and as we are in STREAM
mode it leads to 'write()' will be called once.
Deignations here and below:
H2G - host to guest transmission. Server is host, client is guest.
G2H - guest to host transmission. Server is guest, client is host.
C - copy mode.
ZC - zerocopy mode.
ZU - zerocopy with user poll mode. This mode is removed from test at
this moment, because I need to support SO_RCVLOWAT logic in it.
So, rows corresponds to dequeue mode, while columns show number of
Maybe it would be better to label the rows, I guess the first one is C
and the second one ZC?
Maybe it would be better to report Gbps so if we change the amount of
data exchanged, we always have a way to compare.
bytes
to dequeue in each mode. Each cell contains several values in the next
format:
*------------*
| A / B |
| C |
| D |
*------------*
A - number of seconds which server spent in tx loop.
B - number of seconds which client spent in rx loop.
C - number of seconds which client spent in rx loop, but except 'poll()'
system call(e.g. only in dequeue system calls).
D - Average number of ns for each POLLIN wake up(in other words
it is average value for C).
I see only 3 values in the cells, I missed which one is C and which one
is D.
G2H:
#0 #1 #2 #3 #4 #5
*----*---------*---------*---------*---------*---------*---------*
| | | | | | | |
| | 4Kb | 16Kb | 64Kb | 128Kb | 256Kb | 512Kb |
| | | | | | | |
*----*---------*---------*---------*---------*---------*---------*
| | 2.3/2.4 |2.48/2.53|2.34/2.38|2.73/2.76|2.65/2.68|3.26/3.35|
| | 7039 | 15074 | 34975 | 89938 | 162384 | 438278 |
*----*---------*---------*---------*---------*---------*---------*
| |2.37/2.42|2.36/1.96|2.36/2.42|2.43/2.43|2.42/2.47|2.42/2.46|
| | 13598 | 15821 | 29574 | 43265 | 71771 | 150927 |
*----*---------*---------*---------*---------*---------*---------*
H2G:
#0 #1 #2 #3 #4 #5
*----*---------*---------*---------*---------*---------*---------*
| | | | | | | |
| | 4Kb | 16Kb | 64Kb | 128Kb | 256Kb | 512Kb |
| | | | | | | |
*----*---------*---------*---------*---------*---------*---------*
| | 1.5/5.3 |1.55/5.00|1.60/5.00|1.65/5.00|1.65/5.00|1.74/5.00|
| | 17145 | 24172 | 72650 | 143496 | 295960 | 674146 |
*----*---------*---------*---------*---------*---------*---------*
| |1.10/6.21|1.10/6.00|1.10/5.48|1.10/5.38|1.10/5.35|1.10/5.35|
| | 41855 | 46339 | 71988 | 106000 | 153064 | 242036 |
*----*---------*---------*---------*---------*---------*---------*
Here are my thoughts about these numbers(most obvious):
1) Let's check C and D values. We see, that zerocopy dequeue is faster
on big buffers(in G2H it starts from 64Kb, in H2g - from 128Kb). I
think this is main result of this test(at this moment), that shows
performance difference between copy and zerocopy).
Yes, I think this is expected.
2) In G2H mode both server and client spend almost same time in rx/tx
loops(see A / B in G2H table) - it looks good. In H2G mode, there is
significant difference between server and client. I think there are
some side effects which produces such effect(continue to analyze).
Perhaps a different cost to notify the receiver? I think it's better to
talk about transmitter and receiver, instead of server and client, I
think it's confusing.
3) Let's check C value. We can see, that G2H is always faster that H2G.
In both copy and zerocopy mode.
This is expected because the guest queues buffers up to 64K entirely,
while the host has to split packets into the guest's preallocated
buffers, which are 4K.
4) Another interesting thing could be seen for example in H2G table,
row #0, col #4 (case for 256Kb). Number of seconds in zerocopy mode
is smaller than in copy mode(1.25 vs 2.42), but whole rx loop was
I see 1.65 vs 1.10, are these the same data, or am I looking at it
wrong?
faster in copy mode(5 seconds vs 5.35 seconds). E.g. if we account
time spent in 'poll()', copy mode looks faster(even it spends more
time in 'read()' than zerocopy loop in 'getsockopt()' + 'madvise()').
I think, it is also not obvious effect.
So, according 1), it is better to use zerocopy, if You need to process
big buffers, with small rx waitings(for example it nay be video stream).
In other cases - it is better to use classic copy way, as it will be
more lightweight.
All tests were performed on x86 board with 4-core Celeron N2930 CPU(of
course it is, not a mainframe, but better than test with nested guest)
and 8Gb of RAM.
Anyway, this is not final version, and I will continue to improve both
kernel logic and performance tests.
Great work so far!
Maybe to avoid having to rebase everything later, it's already
worthwhile to start using Bobby's patch with skbuff.
SUGGESTIONS
1) I'm also working on MSG_ZEROCOPY support for virtio/vsock. May be I
can merge both patches into single one?
This is already very big, so I don't know if it's worth breaking into a
preparation series and then a series that adds both.
2) This version works with single page headers. May be I can add new
'getsockopt()' feature to allow to use multiple pages for headers.
What would be the benefit?
A small suggestion, run checkpatch with --strict, because there are
several warnings that would be better resolved.
I'll take a quick look at the patches, but I'd rather do a more detailed
review with skbuffs.
Thanks,
Stefano