[PATCH v4 00/21] add "windowsize" (RFC 7440) support for tftp

Enrico Scholz enrico.scholz at sigma-chemnitz.de
Tue Aug 30 00:37:55 PDT 2022


The tftp "windowsize" greatly improves the performance of tftp
transfers.  This patchset adds support for it.

The first two patches are a little bit unrelated and enhance the 'cp
-v' output by giving information about the transfer speed.  They can
be dropped if they are unwanted.

I tested the function with an iMX8MP platform in three environments:

  - at home over OpenVPN on an ADSL 50 line  -->  27x speedup
  - 1 Gb/s connection --> 9x speedup
  - connection over 100 Mb/s switch  -->  4x speedup

In the test, I downloaded variable sized files which were filled from
/dev/urandom.  E.g.

| :/ global tftp.windowsize=128
| :/ cp -v /mnt/tftp/data-100MiB /tmp/data && sha1sum /tmp/data
|         [################################################################] 104857600 bytes, 98550375 bytes/s

For slow connection speeds, smaller files (1MiB, 4 MiB + 20 MiB) were
used.

The numbers (bytes/s) are

 | windowsize | VPN       | 1 Gb/s     | 100 Mb/s   |
 |------------|-----------|------------|------------|
 | 128        | 3.869.284 | 98.643.085 | 11.434.852 |
 |  64        | 3.863.581 | 98.550.375 | 11.434.852 |
 |  48        | 3.431.580 | 94.211.680 | 11.275.010 |
 |  32        | 2.835.129 | 85.250.081 | 10.985.605 |
 |  24        | 2.344.858 | 77.787.537 | 10.765.667 |
 |  16        | 1.734.186 | 67.519.381 | 10.210.087 |
 |  12        | 1.403.340 | 61.972.576 |  9.915.612 |
 |   8        | 1.002.462 | 50.852.376 |  9.016.130 |
 |   6        |   775.573 | 42.781.558 |  8.422.297 |
 |   4        |   547.845 | 32.066.544 |  6.835.567 |
 |   3        |   412.987 | 26.526.081 |  6.322.435 |
 |   2        |   280.987 | 19.120.641 |  5.494.241 |
 |   1        |   141.699 | 10.431.516 |  2.967.224 |
 |------------|-----------|------------|------------|
 | unpatched  |   140.587 | 10.553.301 |  2.978.063 |


Patchset has been tested with

| for i in data-0 data-100B data-1KiB data-1432B data-64KiB data-1MiB data-4MiB; do
|     tftp    "$i"
|     tftp -p "$i"
| done

against tftp servers with and without rfc 2747 support (OACK).


The window size related parts of the patchset (with deactivated
selftest) increase the barebox binary size by

| add/remove: 6/0 grow/shrink: 7/2 up/down: 1572/-32 (1540)
| Function                                     old     new   delta
| tftp_handler                                 756    1324    +568
| tftp_allocate_transfer                         -     196    +196
| tftp_put_data                                  -     184    +184
| tftp_window_cache_remove                       -     124    +124
| tftp_window_cache_get_pos                      -     120    +120
| tftp_send                                    296     412    +116
| tftp_do_open                                 428     512     +84
| tftp_states                                    -      72     +72
| tftp_do_close                                260     312     +52
| tftp_init                                     16      60     +44
| tftp_open                                     64      68      +4
| tftp_lookup                                  136     140      +4
| g_tftp_window_size                             -       4      +4
| tftp_read                                    180     164     -16
| tftp_poll                                    180     164     -16
| Total: Before=629556, After=631096, chg +0.24%

Turning of the datagram cache (CONFIG_FS_TFTP_REORDER_CACHE_SIZE=0)
reduces the overhead to

| add/remove: 3/0 grow/shrink: 6/2 up/down: 808/-32 (776)
| Function                                     old     new   delta
| tftp_handler                                 756    1092    +336
| tftp_allocate_transfer                         -     144    +144
| tftp_send                                    296     412    +116
| tftp_do_open                                 428     512     +84
| tftp_states                                    -      72     +72
| tftp_init                                     16      60     +44
| tftp_open                                     64      68      +4
| tftp_lookup                                  136     140      +4
| g_tftp_window_size                             -       4      +4
| tftp_read                                    180     164     -16
| tftp_poll                                    180     164     -16
| Total: Before=629556, After=630332, chg +0.12%

Restoring the old behaviour by CONFIG_FS_TFTP_MAX_WINDOW_SIZE=1 shows
an overhead of

| add/remove: 3/0 grow/shrink: 6/2 up/down: 720/-32 (688)
| Function                                     old     new   delta
| tftp_handler                                 756    1088    +332
| tftp_allocate_transfer                         -     144    +144
| tftp_do_open                                 428     512     +84
| tftp_states                                    -      72     +72
| tftp_init                                     16      60     +44
| tftp_send                                    296     328     +32
| tftp_open                                     64      68      +4
| tftp_lookup                                  136     140      +4
| g_tftp_window_size                             -       4      +4
| tftp_read                                    180     164     -16
| tftp_poll                                    180     164     -16
| Total: Before=629556, After=630244, chg +0.11%


Enrico Scholz (21):
  tftp: add some 'const' annotations
  tftp: allow to change tftp port
  cmd:tftp: add '-P' option to set tftp server port number
  tftp: do not set 'tsize' in WRQ requests
  tftp: assign 'priv->block' later in WRQ
  tftp: minor refactoring of RRQ/WRQ packet generation code
  tftp: replace hardcoded blksize by global constant
  tftp: remove sanity check of first block
  tftp: add debug_assert() macro
  tftp: allocate buffers and fifo dynamically
  tftp: add sanity check for OACK response
  tftp: record whether tftp file is opened for lookup operation only
  tftp: reduce block size on lookup requests
  tftp: refactor data processing
  tftp: detect out-of-memory situations
  tftp: implement 'windowsize' (RFC 7440) support
  tftp: do not use 'priv->block' for RRQ
  tftp: reorder tftp packets
  tftp: add selftest
  tftp: accept OACK + DATA datagrams only in certain states
  tftp: add some documentation about windowsize support

 Documentation/filesystems/tftp.rst |  38 ++
 commands/tftp.c                    |  22 +-
 fs/Kconfig                         |  36 ++
 fs/tftp-selftest.h                 |  56 +++
 fs/tftp.c                          | 763 +++++++++++++++++++++++++----
 test/self/Kconfig                  |   7 +
 6 files changed, 824 insertions(+), 98 deletions(-)
 create mode 100644 fs/tftp-selftest.h

---
v3 -> v4
  - fix operation with non rfc 2347 servers
  - do not send 'tsize' in WRQ requests
  - add more sanity checks
  - add some documentation

v2 -> v3
  - use "port=XX" mount options instead of global 'tftp.port' variable
  - allocate fifo and send buffer dynamically based on block- and
    window size of the transfer.  Do not use fixed constants anymore
  - rewritten cache code; use bitmap based functions with O(1)
    complexity instead of iterating over (small) arrays
  - unittest for cache functions
  - add information about binary sizes

v1 -> v2
  - fixes for non rfc7440 servers

-- 
2.37.2




More information about the barebox mailing list