Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2019 (Shanghai), 14 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Preventing DMA Attacks from Thunderbolt Devices - Lu Baolu, Intel

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Preventing DMA Attacks from Thunderbolt Devices - Lu Baolu, Intel

Thunderbolt, shipped on MacOS and Windows platforms, allows the connection of external peripherals directly through the PCIe bus to a computer. Although silicon vendors have DMA remapping technologies to prevent DMA attacks, systems are still vulnerable to DMA attacks from malicious peripherals connected through the thunderbolt ports. This session will present the security issues we have identified with Linux, and solutions we have developed.

https://sched.co/NruQ

A

Okay, I've gotta know: I am bubalu for Intel and I'm, basically working on the Linux community and I'm a developer, and today, I want to talk about like to take this opportunity to share with you some of the issues that encountered previously and I'm still working on solving them.

B

A

Is what I'm going to talk about the sister gender today? First, a brief intro and the second will be followed by device classification in security, environment and the third is vulnerabilities that we have discovered. Then the fixed.

B

A

All, rather with GNA and.

B

A

Thoughts too sure and our to do list and finally, a conclusion and some references first before this session I want you a statement. All these wonderfully vulnerabilities discovered. Then the slides is not.

A

Specific specified to several one or several skin vendor or platform, so I'm, not talking about on in South, Pacific, repos and vulnerabilities, which do helping the communities to fix all this vulnerabilities. All these platforms will benefit from our work and.

B

A

All of these I'd like to talk about very tiny story around one play a thousand years ago by the end of from Eastern Han Dynasty. We had this main part of determining the face of the China North in China. Guan do battle sounds how defeated the UN Chao major wallowed. Turning on a new chapter of the history and because one of the advisor of you and Xiao Ming advisory, mr.

A

Xu, you betrayed you and channel and give gists and vital information to sell, sell and tell us how burned are the wagons of Yong Chun in which are, and then he defeated on you and child, and so this is what we have learned information securities so so important. This.

C

A

The gist we have drawn from that one point: a thousand years ago battle and unfortunately, to today.

B

A

All this we've learned still, we have problems of information, security.

C

A

Still face multiple challenges: more than computers and devices, we have a DMA and axes, and that leads to some.

A

Malicious devices that they want to have access to your memories and leak, your information and still your services. So this is DMA attack.

B

A

Attack is basically a long discussed topic. Little just tell you why we're still focusing on these parts, let's.

B

Do focusing on this issue.

A

And all the signal vendors, including.

B

A

Platforms, they are providing other on. Our I owe em mechanism to protect our device against the other attack. So with IOM a you, are we totally risk-free? No still, we have a lot to do.

C

A

Sharing this is basic concepts. You.

B

A

What are the devices can be trusted? What are the devices that we cannot trust? Well, the trusted trust.

B

With on device.

A

Will follow the order of our drivers and all the DMA behaviors are limited in the buffer and they will not give it. They will not access other device that out.

C

A

And if the drivers are not allowing device to do DMA, then they will not do it and if DMA fail and if it generates the issue, they will report following the same routine to the driver. So these are trustworthy device. This is a trusted. This is what we call trusted device and the untrusted device are the device and that's not equipped with this species so for the external third-party device.

A

How much can we trust it and that must be different from the device that's used internally. So when we are discussed about this, the performance and security of information is different. Sometimes you have to sacrifice part of the security to improve its performance and sometimes vice versa. Well, you had to be more secure at this at the stake of performance so for the trustworthy device we'll have to up and optimize the efficiency of data transmission and for the untrustworthy device, we have to fend it off any risks. So after all, this information has been presented.

A

How can we classify this devices Linux kernel and luckily my Linux communities? They.

B

Have their own classification.

A

For devices, so they think that all the devices are trustworthy and we think that some advice might cost and security issues. We have two categories for them: one trusted device secure.

C

A

And, generally speaking, if this is for your vendor or we put it in the cabinet, then it's trustworthy and if we do hot plug to it or if we have a PICU to it, fire wire, for example, then we think it's untrustworthy, not.

C

C

A

A joke in the community, if you're a device.

C

A

C

A

Will have to open your.

C

A

Manually with a port.

B

A

They have that chance to open the cabinet, they might as well smashed it with a hammer. So we believe that if we put the device internally into a cabinet, it will be very.

B

Safe and it's trustworthy so.

C

B

Submit the pash.

A

B

C

A

Have a total we have interface.

C

A

Pcie slots they will work for it on all the untrustworthy device from internally I would run the tests see the device, whether it's trustworthy or not. So these are trustworthy device. We will identify it and.

B

A

We will have a per device, testing and screening and finally,.

B

A

Trust, an untrustworthy device will have to be determined by the customers, see which are the safe and secure devices which are not so these are what we're working on and, first of all, abilities. Io mu is designed to protect the device fend it off. Dma attack, so.

C

A

A

I'm, citing PhD research and thesis from Cambridge University.

B

A

Discovered that the FreeBSD Mac OS.

B

A

Platforms, only Mac OS can actually fend off all the attacks and none for the other, not for the other platforms and for Windows for the latest version. Enterprise version windows.

B

A

Have to look through relevant documents to do this, if you're using three bsd and quoting a 4 and 16 point o4 and Red Hat Enterprise Linux 7 point 1: these are the.

B

A

Was switched off by default and in 2000 8080 we have run a read research if we find an untrustworthy device, the system will turn on the eyeliner you first screen manually and this a prevention and second vulnerability.

B

A

If you are familiar with.

B

A

Dx- and this is called address- translation service featured defined by PCIe. If we support 80s, we will that's a little request to Iowa. You.

B

A

Send a application, 80s translation request to Iowa a you. They will bypass the device used. The translated address for DMA.

B

A

I all may you upon seeing this translation request using the translated address. They will bypass this request.

B

A

A

C

The efficiency of DMA.

A

So this is pre translated, not translated, really site so also.

C

A

Means if you're malicious device in a support IDs as a disguised and you have access to any memory in the system than your privilege device actually. So this is a potential risks.

B

A

We have short term long.

B

Term solutions.

A

For short term solution for our trusted device, would it will always disable IDs with dead? This part.

C

B

A

A

Released it in version 5.1 and for long term solution. We will security as support and improve the tight sufficiency and.

C

A

We have PCIe ECM.

A

Try to solve this problem for.

C

A

Page, this is the third vulnerability. Oh.

C

A

Have I wanna you window, it is pretty similar to DB you a menu. It's.

A

Cutting a virtual space into into different pages, each page might be larger than 4k. So one page is this:.

B

A

Smallest units and all this device.

C

B

A

Given bite of one page.

C

A

Blue part is the driver buffer assigned to the system, managed device.

C

A

Acces to the whole page and the red part I not meant for the.

B

A

Accessing but it was exposed to the memory and then all the data and cold in the memory is not safe and nor any longer you have the risk of information leakage from here.

B

A

Is our solution in the upper graph, the device water clean hazard this buffer and in our solution we add the past page, so the device visit not the final buffer but the best page to the right. So I don't know about your buffer mechanism.

A

The blue buffer is different from the red buffer, so all the device see are the pages to the right. That includes the blue parts that drive a buffer.

A

The the green parts.

B

A

Basically, the best page are not yet realistic and we're still testing it and.

A

If we're releasing version 5, then we will do wireline and it will be almost done we're trying to solve this proverb by.

B

A

Means so we're using past buffers for software, then for hardware we are also finding ways, driving paths and RI is pretty low, and the overhead, if you're using fast paid overhead, is too high and.

A

We didn't have a formal proposal, yet we're still considering different we're still weighing among possible solutions and the fourth vulnerability is strict mode. If you're not an expert of our and then you, then you probably don't know about it, but.

B

A

Pretty simple, actually so.

B

A

Doing DMA we have them do two parts of their. The array starts the.

B

First part includes.

A

Several ports, for example, not buffer, DMA transfer and buffer, not IOT, ILP invalidation and the DNA and.

C

A

A

The puffer our map and IOT out the invalidation cannoli have one at the same time.

B

A

One at a time- and we have this proposal.

A

Then they propose that we shift the report and break it down to different tasks.

C

A

Have tasks you.

B

A

We will made at the DMA like this first buffer map, DMA transfer and them dearly and the task queue.

A

Will focus on the implementation? It's like a lazy mold. That's what we call it.

C

A

Now the lazy mode is not done simultaneously.

C

A

There's one issue that.

B

A

A very long window, your devices, do you have access to an internal.

B

A

C

A

Location probably has been transferred to other device by kernel, and here is a potential risk threats.

B

A

Dma is now finished, but your map still enable the cache and your device can still visit the cache and the kernel memory allocation system probably has already assigned this memory to other device so for all trusted Oh serious. We will always use this trick mode. We always always apply this trick more.

B

A

Means we're using the upper taking the upper path and guaranteed it's secure, so we will divide it into two groups. The trusted devise an untrusted vise.

A

We're doing this to guarantee that maximum.

B

D

This was darling April of 1991 and has been intended, so we have discussed before security the course and how to solve.

B

D

Many people are doing this research on is the similarity with the DMA attack the security issues. When it's not just to these four types of security issues, we have more issues than that. I am developing I. Am you.

B

D

Can we just jump out.

B

Of the box I, am you and.

C

D

If the time window.

D

I mean your window is not in the import. When the structure is not English initialized, it cannot be upgraded, but each platform has its own procedures to deal.

B

D

It humidity, we have the protective memory region, it ensures.

B

D

The data exported and until the memory is enabled the internal memory is kept secure. I.

C

B

Of the drivers.

C

D

D

Travers encounter an unrecoverable error.

B

D

Trigger recoverable errors.

B

D

They think that device is trusted when the device acts abnormally. They think the system is breaking there's an issue with assistant, so this pulled the system into the unpredictable status. We don't think this is a good method.

C

D

Malicious device, if you encountered the issue as to how can we deal with it, so you need to improve the robustness of the driver.

B

D

The third one being the OS.

D

Well, report errors.

D

But they encounter Linux other wrong messages. What captain D messages alone or in.

B

C

D

Don't have this automatic mechanism to tell that the driver that to make it a responder yeah is it doesn't stop the device or function reset the function? There's a gap here so.

B

The last a should be back.

D

When the so how to talk to achieve our identify devices, so it read, is manufacturer and device ID.

D

Device IT it's a publicly available on the internet.

B

D

Example, the device is made by one person specifically facture, so it provides the convenience of others and malicious man attacker. It can copy any type of the device and to attack it. So I haven't mentioned a doctor.

B

D

University being FDA.

B

D

Writes the code and if he just made the malicious, the.

D

Devices and he castle were seeing a lot of issues.

C

Which, as malicious, was supposed you.

D

Are the developers and you will develop at all types of device driver.

B

D

You are writing component. A lot. Do not use panic, for example. If then, this is not your good.

B

D

You just opposes this device is trusted, but maybe it's not it's untrusted. It's so suppose it's untrusted in the world so avoid using panic. You can use a warm.

C

D

B

D

You can you can lock all this information in the kernel lock and go down.

B

D

When you are using the platform which they use, the page aligned buffer.

B

D

As to bypass the prospect buffer mechanism,.

B

When you're doing a DMA, please.

D

Specify the crack with EMI traction and is it from the kernel to memory or from the memory to the device so in in conclusion, on synchronous, silica and upside for.

C

D

Provide a barrier go technologists to protect the system that guess DMA attacks.

B

D

Os, they relax the controller, mat DMA. So why? Because they think the device is trusted. They are more focused on improving the data transmission efficiency.

C

D

B

B

D

You know classified devices into different the trust levels. Currently we have two levels, but in the future we can expand it and segment it into more levels. After we have segment 8h and for the trusted devices, we will keep optimizations for the efficient attina.

B

D

For those untrusted devices we focus on is a security and another thing being that we were when we are getting. How do we respect.

D

We need to pay attention to the malicious devices to prevent attack.

D

Yes, the first time I have proposed the second one we are still considering it.

B

D

Is the reference.

B

D

If you have any questions.

D

Thank you so did have any cross. Yes, so please go.

C

D

D

For the ETS tsk I'll expand on it.

B

D

Is not proposed, my mile, it is proposed by Intel lab, so I have.

C

This link here, if.

D

This is yes, you can see this link here. Yes,.

B

I'll go and check it later.

D

Yes, they have loaded it. Who Internet are you? Can it's made public oh Emily available? So so second, a question said for the poems page: it's an impossible in DMA and API level.

D

Maybe the page is not page scienter.

B

D

The Indian : level we can protect it.

B

D

Development, we don't ask you the drivers to.

B

D

Of travelers haven't done this.

B

D

We are encouraging developers to use the pigeon on the buffer.

B

D

We were sightless buffers.

B

D

Strong basis so.

D

B

D

A question I am ma'am with you, so it is used to be virtualized environment, but the for physically Mirman kind also be applied. Yes, this is a good question, so in community many people propose the idea that I am.

C

D

It was designed to increase the protection of the I, am you, but it is being more used in the virtualized environment for.

B

C

D

I told the girl with you that that I am used may function is to people performed in virtualized environments. It has many other functions for the protection of the CPU further. Yes, you can.

C

Terrify them that.

D

Is for protection when you are designing? Well, it is device. We think it is trusted. We just focus on optimization and efficiency, but.

B

D

You encounter an untrusted device this and a lot of more issues coming up.

B

D

I, don't agree with the idea that I am always just for visualization.

D

So for the Mac OS either I am over you, it's doing very well, so can Windows learn from here, and it was the bondage of way to compare to Linux I. Think I.

B

D

Started this from other peoples, graduation papers, so I'm, a kind of a limit I have yes, I, have added citation here. The references. This is the paper.

B

D

I think the Mac OS you're so fast cause a FireWire. The.

B

D

Proposed by Apple it's the kind of a connecting external devices to the DMA.

B

D

This platform has this function, so that's why they focused on this area. The protection of DMA.

D

Windows because the firewire.

B

D

She would turn into a sent box. Video was the code developed by Intel and apple.

C

D

The Wonder sandbox, which is the 30.0.

C

C

D

We they have unified this interface, because in the past we have this different interfaces. Yes, they combine the two into one thing: it's like a wrong shift. You don't can tell whether it's a sandbox or is a 0.3 version, because in the past you can tell the testing difference like the Linux. It's a pain attention to this issue. Yes, they are focusing on fixing this issue.

B

D

Thank you very much.