TFTP Windowsize Option
Lets a TFTP client and server agree to send several data blocks in a row before waiting for an acknowledgment.
Masotta, P., "TFTP Windowsize Option", RFC 7440, DOI 10.17487/RFC7440, January 2015, <https://www.rfc-editor.org/info/rfc7440>.
@misc{rfc7440,
series = {Request for Comments},
number = 7440,
howpublished = {RFC 7440},
publisher = {RFC Editor},
doi = {10.17487/RFC7440},
url = {https://www.rfc-editor.org/info/rfc7440},
author = {P. Masotta},
title = {{TFTP Windowsize Option}},
pagetotal = 9,
year = 2015,
month = jan,
} RFC text
Contents
Internet Engineering Task Force (IETF) P. Masotta
Request for Comments: 7440 Serva
Category: Standards Track January 2015
ISSN: 2070-1721
TFTP Windowsize Option
Abstract
The "Trivial File Transfer Protocol" (RFC 1350) is a simple,
lockstep, file transfer protocol that allows a client to get or put a
file onto a remote host. One of its primary uses is in the early
stages of nodes booting from a Local Area Network (LAN). TFTP has
been used for this application because it is very simple to
implement. The employment of a lockstep scheme limits throughput
when used on a LAN.
This document describes a TFTP option that allows the client and
server to negotiate a window size of consecutive blocks to send as an
alternative for replacing the single-block lockstep schema. The TFTP
option mechanism employed is described in "TFTP Option Extension"
(RFC 2347).
Status of This Memo
This is an Internet Standards Track document.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Further information on
Internet Standards is available in Section 2 of RFC 5741.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
http://www.rfc-editor.org/info/rfc7440.
Copyright Notice
Copyright (c) 2015 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction ....................................................2
2. Conventions Used in This Document ...............................3
3. Windowsize Option Specification .................................3
4. Traffic Flow and Error Handling .................................4
5. Proof of Concept and Windowsize Evaluation ......................6
6. Congestion and Error Control ....................................7
7. Security Considerations .........................................8
8. References ......................................................9
8.1. Normative References .......................................9
Author's Address ...................................................9
1. Introduction
TFTP is virtually unused for Internet transfers today, TFTP is still
massively used in network boot/installation scenarios including EFI
(Extensible Firmware Interface). TFTP's inherently low transfer rate
has been, so far, partially mitigated by the use of the blocksize
negotiated extension [RFC2348]. Using this method, the original
limitation of 512-byte blocks are, in practice, replaced in Ethernet
environments by blocks no larger than 1468 Bytes to avoid IP block
fragmentation. This strategy produces insufficient results when
transferring big files, for example, the initial ramdisk of Linux
distributions or the PE images used in network installations by
Microsoft WDS/MDT/SCCM. Considering TFTP looks far from extinction
today, this document presents a negotiated extension, under the terms
of the "TFTP Option Extension" [RFC2347], that produces TFTP transfer
rates comparable to those achieved by modern file transfer protocols.
2. Conventions Used in This Document
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
[RFC2119].
In this document, these words will appear with that interpretation
only when in ALL CAPS. Lowercase uses of these words are not to be
interpreted as carrying the significance given in RFC 2119.
3. Windowsize Option Specification
The TFTP Read Request or Write Request packet is modified to include
the windowsize option as follows. Note that all fields except "opc"
MUST be ASCII strings followed by a single-byte NULL character.
2B string 1B string 1B string 1B string 1B
+-------+---~~---+----+---~~---+----+-----~~-----+----+---~~---+----+
| opc |filename| 0 | mode | 0 | windowsize | 0 | #blocks| 0 |
+-------+---~~---+----+---~~---+----+-----~~-----+----+---~~---+----+
opc
The opcode field contains either a 1 for Read Requests or a 2 for
Write Requests, as defined in [RFC1350].
filename
The name of the file to be read or written, as defined in
[RFC1350].
mode
The mode of the file transfer: "netascii", "octet", or "mail", as
defined in [RFC1350].
windowsize
The windowsize option, "windowsize" (case insensitive).
#blocks
The base-10 ASCII string representation of the number of blocks in
a window. The valid values range MUST be between 1 and 65535
blocks, inclusive. The windowsize refers to the number of
consecutive blocks transmitted before stopping and waiting for the
reception of the acknowledgment of the last block transmitted.
For example:
+------+--------+----+-------+----+------------+----+----+----+
|0x0001| foobar |0x00| octet |0x00| windowsize |0x00| 16 |0x00|
+------+--------+----+-------+----+------------+----+----+----+
is a Read Request for the file named "foobar" in octet transfer mode
with a windowsize of 16 blocks (option blocksize is not negotiated in
this example, the default of 512 Bytes per block applies).
If the server is willing to accept the windowsize option, it sends an
Option Acknowledgment (OACK) to the client. The specified value MUST
be less than or equal to the value specified by the client. The
client MUST then either use the size specified in the OACK or send an
ERROR packet, with error code 8, to terminate the transfer.
The rules for determining the final packet are unchanged from
[RFC1350] and [RFC2348].
The reception of a data window with a number of blocks less than the
negotiated windowsize is the final window. If the windowsize is
greater than the amount of data to be transferred, the first window
is the final window.
4. Traffic Flow and Error Handling
The next diagram depicts a section of the traffic flow between the
Data Sender (DSND) and the Data Receiver (DRCV) parties on a generic
windowsize TFTP file transfer.
The DSND MUST cyclically send to the DRCV the agreed windowsize
consecutive data blocks before normally stopping and waiting for the
ACK of the transferred window. The DRCV MUST send to the DSND the
ACK of the last data block of the window in order to confirm a
successful data block window reception.
In the case of an expected ACK not timely reaching the DSND
(timeout), the last received ACK SHALL set the beginning of the next
windowsize data block window to be sent.
In the case of a data block sequence error, the DRCV SHOULD notify
the DSND by sending an ACK corresponding to the last data block
correctly received. The notified DSND SHOULD send a new data block
window whose beginning MUST be set based on the ACK received out of
sequence.
Traffic with windowsize = 1 MUST be equivalent to traffic specified
by [RFC1350].
For normative traffic not specifically addressed in this section,
please refer to [RFC1350] and its updates.
[ DRCV ] <---traffic---> [ DSND ]
ACK# -> <- Data Block# window block#
...
<- |DB n+01| 1
<- |DB n+02| 2
<- |DB n+03| 3
<- |DB n+04| 4
|ACK n+04| ->
<- |DB n+05| 1
Error |<- |DB n+06| 2
<- |DB n+07| 3
|ACK n+05| ->
<- |DB n+06| 1
<- |DB n+07| 2
<- |DB n+08| 3
<- |DB n+09| 4
|ACK n+09| ->
<- |DB n+10| 1
Error |<- |DB n+11| 2
<- |DB n+12| 3
|ACK n+10| ->| Error
<- |DB n+13| 4
- timeout -
<- |DB n+10| 1
<- |DB n+11| 2
<- |DB n+12| 3
<- |DB n+13| 4
|ACK n+13| ->
...
Section of a Windowsize = 4 TFTP Transfer
Including Errors and Error Recovery
5. Proof of Concept and Windowsize Evaluation
Performance tests were run on the prototype implementation using a
variety of windowsizes and a fixed blocksize of 1456 bytes. The
tests were run on a lightly loaded Gigabit Ethernet, between two
Toshiba Tecra Core 2 Duo 2.2 Ghz laptops, in "octet" mode,
transferring a 180 MByte file.
^
|
300 +
Seconds | windowsize | time (s)
| ---------- ------
| x 1 257
250 + 2 131
| 4 76
| 8 54
| 16 42
200 + 32 38
| 64 35
|
|
150 +
|
| x
|
100 +
|
| x
|
50 + x
| x
| x x
|
0 +-//--+-----+-----+-----+-----+-----+-----+-->
1 2 4 8 16 32 64
Windowsize (in Blocks of 1456 Bytes)
Comparatively, the same 180 MB transfer performed over a drive mapped
on Server Message Block (SMB) / Common Internet File System (CIFS) on
the same scenario took 23 seconds.
The comparison of transfer times (without a gateway) between the
standard lockstep schema and the negotiated windowsizes are:
Windowsize | Time Reduction (%)
---------- -----------------
1 -0%
2 -49%
4 -70%
8 -79%
16 -84%
32 -85%
64 -86%
The transfer time decreases with the use of a windowed schema. The
reason for the reduction in time is the reduction in the number of
the required synchronous acknowledgements exchanged.
The choice of appropriate windowsize values on a particular scenario
depends on the underlying networking technology and topology, and
likely other factors as well. Operators SHOULD test various values
and SHOULD be conservative when selecting a windowsize value because
as the former table and chart shows, there is a point where the
benefit of continuing to increase the windowsize is subject to
diminishing returns.
6. Congestion and Error Control
From a congestion control (CC) standpoint, the number of blocks in a
window does not pose an intrinsic threat to the ability of
intermediate devices to signal congestion through drops. The rate at
which TFTP UDP datagrams are sent SHOULD follow the CC guidelines in
Section 3.1 of [RFC5405].
From an error control standpoint, while [RFC1350] and subsequent
updates do not specify a circuit breaker (CB), existing
implementations have always chosen to fail under certain
circumstances. Implementations SHOULD always set a maximum number of
retries for datagram retransmissions, imposing an appropriate
threshold on error recovery attempts, after which a transfer SHOULD
always be aborted to prevent pathological retransmission conditions.
An implementation example scaled for an Ethernet environment
(1 Gbit/s, MTU=1500) would be to set:
windowsize = 8
blksize = 1456
maximum retransmission attempts per block/window = 6
timeout between retransmissions = 1 S
minimum inter-packet delay = 80 uS
Implementations might well choose other values based on expected
and/or tested operating conditions.
7. Security Considerations
TFTP includes no login or access control mechanisms. Care must be
taken when using TFTP for file transfers where authentication, access
control, confidentiality, or integrity checking are needed. Note
that those security services could be supplied above or below the
layer at which TFTP runs. Care must also be taken in the rights
granted to a TFTP server process so as not to violate the security of
the server's file system. TFTP is often installed with controls such
that only files that have public read access are available via TFTP.
Also listing, deleting, renaming, and writing files via TFTP are
typically disallowed. TFTP file transfers are NOT RECOMMENDED where
the inherent protocol limitations could raise insurmountable
liability concerns.
TFTP includes no protection against an on-path attacker; care must be
taken in controlling windowsize values according to data sender, data
receiver, and network environment capabilities. TFTP service is
frequently associated with bootstrap and initial provisioning
activities; servers in such an environment are in a position to
impose device or network specific throughput limitations as
appropriate.
This document does not add any security controls to TFTP; however,
the specified extension does not pose additional security risks
either.
8. References
8.1. Normative References
[RFC1350] Sollins, K., "The TFTP Protocol (Revision 2)", STD 33,
RFC 1350, July 1992,
<http://www.rfc-editor.org/info/rfc1350>.
[RFC2347] Malkin, G. and A. Harkin, "TFTP Option Extension", RFC
2347, May 1998, <http://www.rfc-editor.org/info/rfc2347>.
[RFC2348] Malkin, G. and A. Harkin, "TFTP Blocksize Option", RFC
2348, May 1998, <http://www.rfc-editor.org/info/rfc2348>.
[RFC5405] Eggert, L. and G. Fairhurst, "Unicast UDP Usage
Guidelines for Application Designers", BCP 145, RFC 5405,
November 2008, <http://www.rfc-editor.org/info/rfc5405>.
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119, March 1997,
<http://www.rfc-editor.org/info/rfc2119>.
Author's Address
Patrick Masotta
Serva
EMail: patrick.masotta.ietf@vercot.com
URI: http://www.vercot.com/~serva/
AI ELI5
On this page
§1Why TFTP Needs a Window
The Trivial File Transfer Protocol (TFTP, RFC 1350) is a lockstep protocol: it sends one block and waits for its acknowledgment before sending the next. The RFC notes that TFTP is still widely used when computers boot or install over a network, including with EFI (Extensible Firmware Interface).
The blocksize option (RFC 2348) already helped. In practice, on Ethernet, blocks are no larger than 1468 bytes to avoid IP fragmentation. That is still not enough for big files such as the initial ramdisk of a Linux distribution or the PE images used in Microsoft network installations.
This RFC adds a negotiated option, built on the TFTP option extension (RFC 2347), that aims at transfer rates comparable to modern file transfer protocols.
Waiting after every single page is slow, even if the pages are bigger. Sending a few pages at a time before waiting is a better fix.
§3The Windowsize Option in a Request
The option rides in the normal Read Request (opcode 1) or Write Request (opcode 2). After the filename and mode come two more strings: windowsize (case insensitive) and the number of blocks as a base-10 ASCII string. Every field except the opcode is an ASCII string ending in a single NULL byte.
The value MUST be between 1 and 65535 blocks, inclusive. The RFC's example is a Read Request for the file foobar in octet mode with a windowsize of 16, using the default 512-byte block size.
If the server accepts the option, it answers with an Option Acknowledgment (OACK). The value in the OACK MUST be less than or equal to the client's value. The client then MUST either use that value or end the transfer with an ERROR packet with error code 8.
The asker writes on the request slip how many pages it can take at once. The sender may say a smaller number, but never a bigger one.
2B string 1B string 1B string 1B string 1B
| opc |filename| 0 | mode | 0 | windowsize | 0 | #blocks| 0 |
|0x0001| foobar |0x00| octet |0x00| windowsize |0x00| 16 |0x00|§4Traffic Flow and Error Recovery
The data sender (DSND) MUST send the agreed number of consecutive data blocks, then normally stop and wait. The data receiver (DRCV) MUST acknowledge the last block of the window to confirm that the whole window arrived.
Recovery is driven by the last good ACK:
- If an expected ACK does not arrive in time, the last ACK received sets where the next window begins.
- If the receiver sees a block out of sequence, it SHOULD send an ACK for the last block it received correctly. The sender SHOULD then start a new window based on that ACK.
A window shorter than the negotiated size is the final window. If the windowsize exceeds the data to send, the first window is the final one. Traffic with windowsize 1 MUST be equivalent to RFC 1350 traffic, and the final-packet rules are unchanged.
If one page in the stack goes missing, the receiver says which page it last got properly, and the sender starts the next stack from there.
- Data blocks n+1 to n+4A full window of four blocks.
- ACK n+4Acknowledges the last block of the window.
- Data blocks n+5 to n+7Block n+6 suffers an error on the way.
- ACK n+5Last block received correctly.
- Data blocks n+6 to n+9New window starts after the ACK.
§5Measured Speedups
The author tested a prototype on a lightly loaded Gigabit Ethernet between two laptops, in octet mode, with a fixed block size of 1456 bytes and a 180 MByte file. Times fell from 257 seconds with windowsize 1 to 131 (2), 76 (4), 54 (8), 42 (16), 38 (32) and 35 (64). The same transfer over a mapped SMB/CIFS drive took 23 seconds.
As a time reduction, that is 49% at windowsize 2, 70% at 4, 79% at 8, 84% at 16, 85% at 32 and 86% at 64. The RFC credits the gain to fewer synchronous acknowledgments.
Because returns diminish, operators SHOULD test various values and SHOULD be conservative. The best value depends on the network technology and topology.
Sending a few pages at once made the job much faster, but sending a huge stack only helped a little more. A medium stack is the sweet spot.
§6Congestion and Retry Limits
A big window is not in itself a threat to congestion signaling through drops, but the rate at which TFTP UDP datagrams are sent SHOULD follow the congestion control guidelines in Section 3.1 of RFC 5405.
Implementations SHOULD also set a maximum number of retransmission retries and abort the transfer after that, to prevent pathological retransmission. The RFC gives one example for 1 Gbit/s Ethernet with MTU 1500: windowsize 8, blksize 1456, 6 retransmission attempts per block or window, a 1 second timeout and a minimum inter-packet delay of 80 microseconds. Other values may suit other conditions.
Even when you send a stack at once, you should not flood the road, and you should give up politely if the other side never answers.
§7Security Considerations
TFTP has no login or access control. Where authentication, access control, confidentiality or integrity checking are needed, those services must come from a layer above or below TFTP. Servers are often set up to serve only publicly readable files, and listing, deleting, renaming and writing are typically disallowed.
TFTP also has no protection against an on-path attacker, so windowsize values should be chosen according to what the sender, receiver and network can handle. The RFC states that this extension adds no security controls and poses no additional security risks.
TFTP never checks who is asking, and this option does not change that. It also does not make things any riskier.