1 files changed, 117 insertions, 0 deletions
diff --git a/pkg/sentry/fsimpl/ext/README.md b/pkg/sentry/fsimpl/ext/README.md
new file mode 100644
index 000000000..af00cfda8
--- /dev/null
+++ b/pkg/sentry/fsimpl/ext/README.md
@@ -0,0 +1,117 @@
+## EXT(2/3/4) File System
+
+This is a filesystem driver which supports ext2, ext3 and ext4 filesystems.
+Linux has specialized drivers for each variant but none which supports all. This
+library takes advantage of ext's backward compatibility and understands the
+internal organization of on-disk structures to support all variants.
+
+This driver implementation diverges from the Linux implementations in being more
+forgiving about versioning. For instance, if a filesystem contains both extent
+based inodes and classical block map based inodes, this driver will not complain
+and interpret them both correctly. While in Linux this would be an issue. This
+blurs the line between the three ext fs variants.
+
+Ext2 is considered deprecated as of Red Hat Enterprise Linux 7, and ext3 has
+been superseded by ext4 by large performance gains. Thus it is recommended to
+upgrade older filesystem images to ext4 using e2fsprogs for better performance.
+
+### Read Only
+
+This driver currently only allows read only operations. A lot of the design
+decisions are based on this feature. There are plans to implement write (the
+process for which is documented in the future work section).
+
+### Performance
+
+One of the biggest wins about this driver is that it directly talks to the
+underlying block device (or whatever persistent storage is being used), instead
+of making expensive RPCs to a gofer.
+
+Another advantage is that ext fs supports fast concurrent reads. Currently the
+device is represented using a `io.ReaderAt` which allows for concurrent reads.
+All reads are directly passed to the device driver which intelligently serves
+the read requests in the optimal order. There is no congestion due to locking
+while reading in the filesystem level.
+
+Reads are optimized further in the way file data is transferred over to user
+memory. Ext fs directly copies over file data from disk into user memory with no
+additional allocations on the way. We can only get faster by preloading file
+data into memory (see future work section).
+
+The internal structures used to represent files, inodes and file descriptors use
+a lot of inheritance. With the level of indirection that an interface adds with
+an internal pointer, it can quickly fragment a structure across memory. As this
+runs along side a full blown kernel (which is memory intensive), having a
+fragmented struct might hurt performance. Hence these internal structures,
+though interfaced, are tightly packed in memory using the same inheritance
+pattern that pkg/sentry/vfs uses. The pkg/sentry/fsimpl/ext/disklayout package
+makes an execption to this pattern for reasons documented in the package.
+
+### Security
+
+This driver also intends to help sandbox the container better by reducing the
+surface of the host kernel that the application touches. It prevents the
+application from exploiting vulnerabilities in the host filesystem driver. All
+`io.ReaderAt.ReadAt()` calls are translated to `pread(2)` which are directly
+passed to the device driver in the kernel. Hence this reduces the surface for
+attack.
+
+The application can not affect any host filesystems other than the one passed
+via block device by the user.
+
+### Future Work
+
+#### Write
+
+To support write operations we would need to modify the block device underneath.
+Currently, the driver does not modify the device at all, not even for updating
+the access times for reads. Modifying the filesystem incorrectly can corrupt it
+and render it unreadable for other correct ext(x) drivers. Hence caution must be
+maintained while modifying metadata structures.
+
+Ext4 specifically is built for performance and has added a lot of complexity as
+to how metadata structures are modified. For instance, files that are organized
+via an extent tree which must be balanced and file data blocks must be placed in
+the same extent as much as possible to increase locality. Such properties must
+be maintained while modifying the tree.
+
+Ext filesystems boast a lot about locality, which plays a big role in them being
+performant. The block allocation algorithm in Linux does a good job in keeping
+related data together. This behavior must be maintained as much as possible,
+else we might end up degrading the filesystem performance over time.
+
+Ext4 also supports a wide variety of features which are specialized for varying
+use cases. Implementing all of them can get difficult very quickly.
+
+Ext(x) checksums all its metadata structures to check for corruption, so
+modification of any metadata struct must correspond with re-checksumming the
+struct. Linux filesystem drivers also order on-disk updates intelligently to not
+corrupt the filesystem and also remain performant. The in-memory metadata
+structures must be kept in sync with what is on disk.
+
+There is also replication of some important structures across the filesystem.
+All replicas must be updated when their original copy is updated. There is also
+provisioning for snapshotting which must be kept in mind, although it should not
+affect this implementation unless we allow users to create filesystem snapshots.
+
+Ext4 also introduced journaling (jbd2). The journal must be updated
+appropriately.
+
+#### Performance
+
+To improve performance we should implement a buffer cache, and optionally, read
+ahead for small files. While doing so we must also keep in mind the memory usage
+and have a reasonable cap on how much file data we want to hold in memory.
+
+#### Features
+
+Our current implementation will work with most ext4 filesystems for readonly
+purposed. However, the following features are not supported yet:
+
+-   Journal
+-   Snapshotting
+-   Extended Attributes
+-   Hash Tree Directories
+-   Meta Block Groups
+-   Multiple Mount Protection
+-   Bigalloc