Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 1 | This document contains a description of portable OpenSSH's random |
| 2 | number collection code. An alternate reading of this text could |
| 3 | well be titled "Why I should pressure my system vendor to supply |
| 4 | /dev/random in their OS". |
| 5 | |
| 6 | Why is this important? OpenSSH depends on good, unpredictable numbers |
| 7 | for generating keys, performing digital signatures and forming |
| 8 | cryptographic challenges. If the random numbers that it uses are |
| 9 | predictable, then the strength of the whole system is compromised. |
| 10 | |
| 11 | A particularly pernicious problem arises with DSA keys (used by the |
| 12 | ssh2 protocol). Performing a DSA signature (which is required for |
| 13 | authentication), entails the use of a 160 bit random number. If an |
| 14 | attacker can predict this number, then they can deduce your *private* |
Damien Miller | df1b645 | 2000-10-25 14:47:35 +1100 | [diff] [blame] | 15 | key and impersonate you or your hosts. |
Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 16 | |
| 17 | If you are using the builtin random number support (configure will |
Damien Miller | df1b645 | 2000-10-25 14:47:35 +1100 | [diff] [blame] | 18 | tell you if this is the case), then read this document in its entirety. |
Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 19 | |
| 20 | Please also request that your OS vendor provides a kernel-based random |
| 21 | number collector (/dev/random) in future versions of your operating |
Damien Miller | df1b645 | 2000-10-25 14:47:35 +1100 | [diff] [blame] | 22 | systems by default. |
Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 23 | |
| 24 | On to the description... |
| 25 | |
| 26 | The portable OpenSSH contains random number collection support for |
| 27 | systems which lack a kernel entropy pool (/dev/random). |
| 28 | |
| 29 | This collector operates by executing the programs listed in |
| 30 | ($etcdir)/ssh_prng_cmds, reading their output and adding it to the |
| 31 | PRNG supplied by OpenSSL (which is hash-based). It also stirs in the |
| 32 | output of several system calls and timings from the execution of the |
| 33 | programs that it runs. |
| 34 | |
| 35 | The ssh_prng_cmds file also specifies a 'rate' for each program. This |
| 36 | represents the number of bits of randomness per byte of output from |
| 37 | the specified program. |
| 38 | |
| 39 | The random number code will also read and save a seed file to |
| 40 | ~/.ssh/prng_seed. This contents of this file are added to the random |
Damien Miller | df1b645 | 2000-10-25 14:47:35 +1100 | [diff] [blame] | 41 | number generator at startup. The goal here is to maintain as much |
| 42 | randomness between sessions as possible. |
Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 43 | |
Damien Miller | df1b645 | 2000-10-25 14:47:35 +1100 | [diff] [blame] | 44 | The entropy collection code has two main problems: |
Damien Miller | 6f9c337 | 2000-10-25 10:06:04 +1100 | [diff] [blame] | 45 | |
| 46 | 1. It is slow. |
| 47 | |
| 48 | Executing each program in the list can take a large amount of time, |
| 49 | especially on slower machines. Additionally some program can take a |
| 50 | disproportionate time to execute. |
| 51 | |
| 52 | This can be tuned by the administrator. To debug the entropy |
| 53 | collection is great detail, turn on full debugging ("ssh -v -v -v" or |
| 54 | "sshd -d -d -d"). This will list each program as it is executed, how |
| 55 | long it took to execute, its exit status and whether and how much data |
| 56 | it generated. You can the find the culprit programs which are causing |
| 57 | the real slow-downs. |
| 58 | |
| 59 | The entropy collector will timeout programs which take too long |
| 60 | to execute, the actual timeout used can be adjusted with the |
| 61 | --with-entropy-timeout configure option. OpenSSH will not try to |
| 62 | re-execute programs which have not been found, have had a non-zero |
| 63 | exit status or have timed out more than a couple of times. |
| 64 | |
| 65 | 2. Estimating the real 'rate' of program outputs is non-trivial |
| 66 | |
| 67 | The shear volume of the task is problematic: there are currently |
| 68 | around 50 commands in the ssh_prng_cmds list, portable OpenSSH |
| 69 | supports at least 12 different OSs. That is already 600 sets of data |
| 70 | to be analysed, without taking into account the numerous differences |
| 71 | between versions of each OS. |
| 72 | |
| 73 | On top of this, the different commands can produce varying amounts of |
| 74 | usable data depending on how busy the machine is, how long it has been |
| 75 | up and various other factors. |
| 76 | |
| 77 | To make matters even more complex, some of the commands are reporting |
| 78 | largely the same data as other commands (eg. the various "ps" calls). |
| 79 | |