Patch the Linux Kernel Without Reboots

evanbro writes "ZDNet is reporting on ksplice, a system for applying patches to the Linux kernel without rebooting. ksplice requires no kernel modifications, just the source, the config files, and a patch. Author Jeff Arnold discusses the system in a technical overview paper (PDF). Ted Ts'o comments, 'Users in the carrier grade linux space have been clamoring for this for a while. If you are a carrier in telephony and don't want downtime, this stuff is pure gold.'" Update: 04/24 10:04 GMT by KD : Tomasz Chmielewsk writes on LKML that the idea seems to be patented by Microsoft.

Comments

Needed that bad?

[MetalliQaZ]

If you are a carrier in telephony, you should have many load-balanced servers that can be taken offline one at a time and restored after patching. They probably would be taken out of the loop for the in-place patching anyway. So who is "clamoring"?

Re:Needed that bad?

[Iphtashu Fitz]

The very fact that there is load balancing means that every server is likely to have active connections going through it. If you currently have connections going through a specific server, you don't want to drop those connections in order to reboot that particular machine. This allows updates to a live machine.

If you have a load balanced environment then you have the ability to redirect new connections away from a given server. Then it's just a matter of waiting for the active connections to terminate before the machine ends up in an idle state where you can safely apply patches offline. I've worked in a number of telephony environments and this was always the way we would patch systems. Stop accepting new connections, wait for existing ones to end, then perform the patch, reboot, verify, and start accepting connections again.

Second, this is telephony, meaning it is the infrastructure on which the internet is based. There's no dns tricks or tcp/ip you can use to send people to a different "server" if that server is the switch connected to your fiber backbone. Basically, there are points in the infrastructure where there are by necessity a single chokepoint.

Any mission critical hardware, switches, routers, servers, etc. should be set up in redundant pairs (or triplets, ...) so that if a hardware failure occurs the remaining hardware can keep the service up. Single points of failure are avoided like the plague in datacenters that require 100% uptime. Part of that is to deal with hardware failures but part is also to provide an ability to perform software/firmware upgrades when necessary. Once again, you migrate all traffic off the system you're upgrading then apply the upgrades offline. Upgrading a kernel, especially, in an online environment, is something virtually any sysadmin would want to avoid if at all possible.

Redundancy is key, and any commercial datacenter will offer it all the way from their connections to the outside world to the connections they provide their customers. Every datacenter used by every company I ever worked for (about 10) offered redundant power and redundant network drops (using HSRP, VRRP, etc) for our equipment. If the datacenter needed to upgrade a router they'd move all traffic off one router so they could upgrade and test it, then move traffic off the other and repeat the process. Similarly if we needed to upgrade our firewalls, switches, etc. we'd fail over to the second redundant device first. In some cases we had bonded interfaces right on the end servers so as long as one path remained active we could power down an entire switch, router, firewall, etc. In other cases we relied on load balancing across servers that were alternately connected to one or another switch.

Re:Needed that bad?

[jelle]

So you take it out of rotation on the load balancer and give it a few minutes to complete all its active connections. Patch/reboot whatever. Bring it back into rotation, and repeat with the other box.

Methods like that usually suck in real-life, because right the day before you want to 'take it out of rotation', a circuit is opened through it that requires five nines (so you can't drop it), and it will remain open for months...

You will end up with 99 boxes waiting to 'get out of rotation' for every
single box that you don't need to update...

Murphy will make sure of that.

Re:Needed that bad?

[Anonymous Coward]

I have internal processing servers that have up times of over 3 years

I've never understood this boasting about uptime. Long uptimes are a bad thing! How do you know a configuration change hasn't rendered one of your startup scripts ineffective? If you have to reboot for some unexpected reason, you could be stuck debugging unrelated problems at very inopportune moments.

You need to schedule regular reboots so that you can test that your servers can start up fine at a moment's notice. Long uptimes are a sign a sysadmin hasn't been doing his job.

Re:Needed that bad?

[mr_mischief]

If you change something in a configuration that requires a change to the startup script, then you also change the startup script.

A patch to the kernel almost never requires changes to startup scripts. They're not talking about adding new functionality with user-space-addressable interfaces with this tool. They're talking about being able to install about 84% of security hotfixes in a hurry outside your scheduled reboots then rebooting on your regular maintenance schedule.

Re:Needed that bad?

[Kookus]

Production systems are not for testing purposes. You want to test rebooting? Do it on a test box.

Re:Needed that bad?

[Kymermosst]

How do you know a configuration change hasn't rendered one of your startup scripts ineffective?

Isn't that what QA systems and effective approaches to change management are supposed to handle?

If I am planning a change, I should discover problems with the startup scripts in QA, not in production, especially if a production reboot is not required to implement the change.

Re:Needed that bad?

[adrianbaugh]

How do you know that your test boxes are configured precisely identically to the production boxes?

dd your production box's system filesystems to another hard drive, put in an identically specced machine, boot that?

Re:Needed that bad?

[mOdQuArK!]

There's a difference between what YOU as an end user consider to be an open connection, and what the telecom equipment consider as a connection.

For all you know, your apparent always-on connection was actually a virtual connection being frequently switched & reswitched over many different real physical connections. That would be a fairly standard architecture for having a network infrastructure which can have components being worked on while data is still flowing through the network.

When the telecom provider is "waiting for active connections to go away" on a particular device only means that all of the virtual connections that are momentarily being switched that device have been successfully switched to another device. It doesn't mean that any of those virtual connections have to be terminated.

Amazing

[cromar]

That is truly amazing tech, right there. It would be interesting to know the security implications of being able to hot-patch the kernel, however.

Re:Amazing

[katz]

Considering that you don't need to prepare the kernel in any way--just execute the program and bang, it's patched--means that someone with root access could slip a rootkit right under your nose (i.e., without the system administrator being aware of this).

- Roey

Re:Amazing

[KeithJM]

someone with root access could slip a rootkit right under your nose Yeah, someone with root access can take control of your server. Oh, wait, they've got root access. They already have control of your server. At some point, you have to just accept that giving someone root access is a security risk.

Wrong way to solve the uptime problem

[Anon E. Muss]

Trying to keep one server up 24/7/365 is a usually mistake. You'll never achieve 100% uptime. A much better idea is to use clustering and distributed computing so your overall system can survive the loss of individual servers.

Re:Wrong way to solve the uptime problem

[Qzukk]

Trust me, that was the first thing they thought of, then the CEO came in and said "Why are you ordering more equipment when we have half of our machines sitting there and doing nothing? We could be doing twice the work/traffic/whatever without paying more money!"

Re:Wrong way to solve the uptime problem

[trybywrench]

Trying to keep one server up 24/7/365 is a usually mistake. You'll never achieve 100% uptime. A much better idea is to use clustering and distributed computing so your overall system can survive the loss of individual servers. People using Linux on BigIron(tm) bank on 24/7/365/25years uptime. When a single server costs hundreds of thousands or millions of dollars you can't afford a spare sitting idle. From day 1 the server needs to be making money and never ever stop. For smaller general purpose servers like you can buy at Dell.com then yeah having a fail-over makes sense.

Re:Wrong way to solve the uptime problem

[Anonymous Coward]

If he knew anything, he wouldn't be the CEO.

Re:Unless it fails.

[Anonymous Coward]

honestly how much downtime are we talking here? 30 seconds? well, think about the fsck that happens after 180 days or 30+ mounts ?

Already been used

[caluml]

There was a kernel exploit recently where someone submitted a patch that modified the running kernel using this technology. It didn't work for me, so I had to resort to patching the .c that was affected - but a lot of people reported that it worked.

Re:Maybe...

[CogDissident]

I thought their working slogan was:

Windows 7, it's not awful like Vista!

The real test...

[hal2814]

Can ksplice be installed without rebooting?

Impressive hack

[EriktheGreen]

For those that haven't read the paper, the technique used is straightforward in concept, but the devil is in the details.

He basically compiles a patched and unpatched kernel with the same compiler, compares the ELF output, and uses that to generate a binary file that corresponds to the change. That gets wrapped in a generic module for use, another module installs it along with JMPs to bypass the old code and use the new, and he performs the checks needed to make sure he can safely install the redirects.

He also has to differentiate real changes from incidental ones (the example given is changing the address of a function - all references to it will change, but they don't really need to be included in the binary diff).

The only human work required is to check whether a patch makes semantic changes to a data structure... whether eg. an unsigned integer variable that was being used as a number is now a packed set of flags - the data declaration is the same, but it's being used differently.

Interesting paper. Also a useful new set of capabilities for any Linux user who can't handle downtime for quarterly patching... worth its weight in gold in some businesses.

Erik

If it's that critical, shouldn't you have two?

[Paul Carver]

I'd rather have at least two of anything important and have statefull failover between them.

If you've got this system that's so critical you can't reboot it for a kernel upgrade, what do you do when the building catches fire or a tanker truck full of toxic waste hops the curb and plows through the wall of your datacenter?

I'd rather have a full second set of anything that critical. It should be in a different state (or country) and have a well designed and frequently used method of seamlessly transferring the load between the two (or more) sites without dropping anything.

If you can't transfer the workload to a location at least a couple hundred miles away without users noticing then you're not in the big league.

And as long as the workload is in another datacenter, what's the big deal about rebooting for a kernel upgrade.

Not only the CEO

[Moraelin]

Not only the CEO. I lived to see even a hardline IT guy (admittedly, one whose goal in life seems to be to be against whatever you want, and to avoid doing any extra work... actually, make that just: any work) argue along the lines of "nooo, you can't have the servers only 60% loaded! It's a waste of valuable hardware! Why, back in my day (of batch jobs on punched cards, presumably) we had the mainframe used at least an average of 95% before asking for an extra server!"

It always irks me to see people just not understand concepts like "peak" vs "average", or "failing over".

- A cluster of, say, 4 machines (small application, really) which are loaded to 90% of capacity, if one dies, the other 3 are now at 120% of capacity each. If you're lucky, it just crawls, if you're unlucky, Java clutches its chest and keels over with an "OutOfMemoryError" or such.

- if you're at 90% most of the time, then fear Monday 9:00 AM, when every single business partner on that B2B application comes to work and opens his browser. Or fear the massive year-end batch jobs, when that machine/cluster sized barely enough to be ready with the normal midnight jobs by 9 AM, so those users can see their new offers and orders in their browsers, now has to do 20 times as much in a burst.

Basically it amazes me how many people just don't seem to get that simple rule of thumb of clusters: you're either getting nearly 100% uptime and nearly guaranteed response times, _or_ you're getting that extra hardware fully used to support a bigger load. Not both. Or not until that cluster is so large that 1-2 servers failing add negligible load to the remaining machines.

Re:Unless it fails.

[Tychon]

A company that I once had dealings with was quite proud of their five nines. The motivation? It cost them $18,000 per second they were down. 30 seconds isn't just 30 seconds sometimes.

No, No, No and No again.

[Anonymous Coward]

As an admin for some -very- high availability systems, load balancers are not a silver bullet. This solution would most apply for running one-node clusters who are using a single machine as a perimeter network device. (ex. firewall) I see lots of these in the racks at our NOC provider.

1. We connect to several load balanced systems and the complexity introduced by load balancers translates to inexplicable down time. No load balancers means a pretty steady diet of the latest and greatest server hardware, but no down time. The a few minutes of down time costs more than the server hardware.

2. High availability translates more roughly into nodes that can fail (ex. power off) and not take the cluster down. This boils down to active-passive application architecture more than just using heartbeat.

As an FYI, PostgreSQL clustering is a killer application for me. Erlang is also great in many ways, but requires application architecture with active-passive node awareness. Which isn't present in things like Yaws, or even my other favorite non-erlang app nginx. Heartbeat is the solution there, but I'd like to see yaws be cluster aware on its own. http://yaws.hyber.org/

Re:No, No, No and No again.

[hab136]

As an admin for some -very- high availability systems, load balancers are not a silver bullet. This solution would most apply for running one-node clusters who are using a single machine as a perimeter network device. (ex. firewall) I see lots of these in the racks at our NOC provider.

1. We connect to several load balanced systems and the complexity introduced by load balancers translates to inexplicable down time. No load balancers means a pretty steady diet of the latest and greatest server hardware, but no down time. The a few minutes of down time costs more than the server hardware.

I spent a decade in perimeter networking at a Fortune 50 US bank. My group didn't do the internal network, just the perimiter, and we still had dozens of network sites and thousands of pieces of equipment. The bank itself has hundreds of thousands of employees, millions of users. Online banking and brokerage are about as high availability as you can get save utilities (power, water, telephony, etc) or military. Seconds of online brokerage downtime equated to millions of dollars lost.

The idea that load balancing introduces inexplicable down time is completely unsupported by my experience.

"One-node clusters" seems like marketing speak for "single point of failure". A cluster by definition is two or more nodes.

Redundant routers, switches, firewalls, the works or you're not high-availability in my opinion. The fact that you're talking about Postgresql instead of Oracle or DB2 on mainframes makes me think that your idea of high availability is different than mine.

Re:In Soviet Russia,

[oodaloop]

"But does it run linux?" That's a joke? I thought that was just one dedicated user who kept asking on every article.