Our research shows that network-based cache side-channel attacks are a realistic threat. Cache attacks have been traditionally used to leak sensitive data on a local setting (e.g., from an attacker-controlled virtual machine to a victim virtual machine that share the CPU cache on a cloud platform). With our attack called NetCAT, we show this threat extends to untrusted clients over the network, which can now leak sensitive data such as keystrokes in a SSH session from remote servers with no local access. The root cause of the vulnerability is a recent Intel feature called DDIO, which grants network devices and other peripherals access to the CPU cache. Originally, intended as a performance optimization in fast networks, we show DDIO has severe security implications, exposing servers in local untrusted networks to remote side-channel attacks.

Increased peripheral performance is causing strain on the memory subsystem of modern processors. For example, available DRAM throughput can no longer sustain the traffic of a modern network card. Scrambling to deliver the promised performance, instead of transferring peripheral data to and from DRAM, modern Intel processors perform I/O operations directly on the Last Level Cache (LLC). While Direct Cache Access (DCA) instead of Direct Memory Access (DMA) is a sensible performance optimization, it is unfortunately implemented without care for security, as the LLC is now shared between the CPU and all the attached devices, including the network card.

In this talk, we present the first security analysis of DDIO. Based on our analysis, we present NetCAT, the first network-based cache attack on the processor’s last-level cache of a remote machine. We show that NetCAT can break confidentiality of a SSH session from a third machine without any malicious software running on the remote server or client. The attacker machine does this by solely sending network packets to the remote server. netcat is also a famous utility that hackers and system administrators use to send information over the network. NetCAT is a pun on being able to read data from the network without cooperation from the other machine on the network. However, we received very mixed reactions on that pun. More details on this in the talk.

The vulnerability was acknowledged by Intel with a bounty and CVE-2019-11184 was assigned to track this issue. The public disclosure was on September 10, 2019.

EnhanceIO driver is based on EnhanceIO SSD caching software product
developed by STEC Inc. EnhanceIO was derived from Facebook's open source
Flashcache project. EnhanceIO uses SSDs as cache devices for
traditional rotating hard disk drives (referred to as source volumes
throughout this document).

EnhanceIO can work with any block device, be it an entire physical
disk, an individual disk partition,  a RAIDed DAS device, a SAN volume,
a device mapper volume or a software RAID (md) device.

The source volume to SSD mapping is a set-associative mapping based on
the source volume sector number with a default set size
(aka associativity) of 512 blocks and a default block size of 4 KB.
Partial cache blocks are not used. The default value of 4 KB is chosen
because it is the common I/O block size of most storage systems.  With
these default values, each cache set is 2 MB (512 * 4 KB).  Therefore,
a 400 GB SSD will have a little less than 200,000 cache sets because a
little space is used for storing the meta data on the SSD.

EnhanceIO supports three caching modes: read-only, write-through, and
write-back and three cache replacement policies: random, FIFO, and LRU.

Read-only caching mode causes EnhanceIO to direct write IO requests only
to HDD. Read IO requests are issued to HDD and the data read from HDD is
stored on SSD. Subsequent Read requests for the same blocks are carried
out from SSD, thus reducing their latency by a substantial amount. 

In Write-through mode - reads are handled similar to Read-only mode.
Write-through mode causes EnhanceIO to write application data to both
HDD and SSD. Subsequent reads of the same data benefit because they can
be served from SSD.

Write-back improves write latency by writing application requested data
only to SSD. This data, referred to as dirty data, is copied later to
HDD asynchronously. Reads are handled similar to Read-only and
Write-through modes.

EnhanceIO can work with any block device, be it an entire physical disk, an individual disk partition, a RAIDed DAS device, a SAN volume, a device mapper volume or a software RAID (md) device.