Slashdot Mirror
Torvalds On Pluggable Security Models
eldavojohn writes "The KernelTrap highlights an interesting discussion on pluggable security models including some commentary by Linus Torvalds. While Torvalds argued against pluggable schedulers, he's all for pluggable security. Other members were voicing concerns with the pluggable nature of the Linux Security Model, but Torvalds put his foot down and said it stays. When asked why his stance was different between schedulers and security, he replied, 'Schedulers can be objectively tested. There's this thing called 'performance,' that can generally be quantified on a load basis. Yes, you can have crazy ideas in both schedulers and security. Yes, you can simplify both for a particular load. Yes, you can make mistakes in both. But the *discussion* on security seems to never get down to real numbers. So the difference between them is simple: one is hard science. The other one is people wanking around with their opinions.'"
He is right, definitely But being theoretically able to measure something doesn't mean it's practical or the the results are always useful.
Ahh, the "when in doubt claim OO is expensive" defense. Please tell me, how long does a modern CPU need to take a branch to an address in a well known fixed memory cell which is guaranteed to be in L1-cache? Do you think it is longer than a conditional branch needed to handle the case single core dual core? Is it longer than the combined times needed to additionally handle the case one CPU-chip two CPU-chips? I don't know, I haven't done the measuring, but I have doubts the first is the slowest as the opcode scheduler should be able to handle the first and especially has the advantage of an always taken jump. We are heading in a parallel future, there are scheduling differences between single core/dual core and single CPU/multiple CPU. Why on earth should the scheduler written for the most complicated case (it has to handle cases like one dual core and two triple cores and one quad core efficiently or it is not the best scheduler, no?) be more efficient than a single core scheduler on a machine with only a single core? Or are the benchmarks "tweaked" so the first is the "right" case to benchmark?
As written by multiple posters, yes, you can get benchmark results for schedulers, but what is the correct benchmark? Is it the maximum throughput model you don't want to have as a desktop box or the minimum waiting time for interactive jobs you don't want on a compute server? And if you need numbers to come up with the best security model, count line numbers, it is about as relevant.
No, Linus' point about schedulers is that to make a pluggable scheduler, you will need to sacrifice performance just to achieve the plug-ability. Linus believes that the most flexible scheduler (i.e. performance, tune-ability) can be discovered at development time with a set of metrics that are defined currently. In which case he feels that the kernel devs can make the "best" choice of scheduler up front. Yes there will be fringe cases, in which case, you have the code, replace/tune/massacre the scheduler to your particular needs.
The security realm however is completely different. For one, the performance requirement does not exist. So the performance penalty that modular architecture brings is largely irrelevant. And since there exist no metrics that can be used to determine whether one security model is better than another without the usage context, a plug-able architecture is the best road to go down to let something that users CAN and WILL want to implement completely differently from one use-case to the next.
-- Humans, because the hardware IS the software.