DLPI and the IP Fastpath

The series so far

If you’re getting started you’ll want to see the previous entries on Project Bardiche:

The illumos Networking Stack

This blog post is going to dive into more detail about what the ‘fastpath’ is in illumos for networking, what it means, and a bit more about how it works. We’ll also go through and cover a bit more information about some of the additions we made as part of this project. Before we go too much further, let’s take another look at the picture of the networking stack from the entry on architecture of vnd:

             +---------+----------+----------+
             | libdlpi |  libvnd  | libsocket|
             +---------+----------+----------+
             |         ·          ·    VFS   |
             |   VFS   ·    VFS   +----------+
             |         ·          |  sockfs  |
             +---------+----------+----------+
             |         |    VND   |    IP    |
             |         +----------+----------+
             |            DLD/DLS            |
             +-------------------------------+
             |              MAC              |
             +-------------------------------+
             |             GLDv3             |
             +-------------------------------+

If you don’t remember what some of these components are, you might want to refresh your memory with the vnd architecture entry. Importantly, almost everything is layered on top of the DLD and DLS modules.

The illumos networking stack comes from a long lineage of technical work done at Sun Microsystems. Initially the networking stack was implemented using STREAMs. STREAMs is a message passing interface where message blocks (mblk_t) are sent around from one module to the next. For example, there are modules for things like arp, tcp/ip, udp, etc. These are chained together and can be seen in mdb using the ::stream dcmd. Here’s an example for my zone development zone:

> ::walk dld_str_cache | ::print dld_str_t ds_rq | ::q2stream | ::stream

+-----------------------+-----------------------+
| 0xffffff0251050690    | 0xffffff0251050598    |
| udp                   | udp                   |
|                       |                       |
| cnt = 0t0             | cnt = 0t0             |
| flg = 0x20204022      | flg = 0x20244032      |
+-----------------------+-----------------------+
            |                       ^
            v                       |
+-----------------------+-----------------------+
| 0xffffff02510523f8    | 0xffffff0251052300    | if: net0
| ip                    | ip                    |
|                       |                       |
| cnt = 0t0             | cnt = 0t0             |
| flg = 0x00004022      | flg = 0x00004032      |
+-----------------------+-----------------------+
            |                       ^
            v                       |
+-----------------------+-----------------------+
| 0xffffff0250eda158    | 0xffffff0250eda060    |
| vnic                  | vnic                  |
|                       |                       |
| cnt = 0t0             | cnt = 0t0             |
| flg = 0x00244062      | flg = 0x00204032      |
+-----------------------+-----------------------+
...

If I sent a udp packet, it would first be processed by the udp streams module, then the ip streams module, and finally make its way to the DLD/DLS layer which is represented by the vnic entry here. The means of this communication was part of the DLPI. DLPI itself defines several different kinds of messages and responses which can be found in the illumos source code here. The general specification is available here, though there’s a lot more to it than is worth reading. In illumos, it’s been distilled down into libdlpi.

Recall from the vnd architecture entry that the way devices and drivers communicate with a datalink is by initially using STREAMS modules and by opening a device in /dev/net/. Each data link in the system is represented by a dls_link_t. When you open a device in /dev/net, you get a dld_str_t which is an instance of a STREAMS device.

The DLPI allows consumers to bind to what they call a SAP or a service attachment point. What this means depends on the kind of data link. In the case of Ethernet, this refers to the ethertype. In other words, a given dld_str_t can be bound to something like IP, ARP, LLDP, etc. If this were something other than Ethernet, then that name space would be different.

For a given data link, only one dld_str_t can be actively bound to a given SAP (ethertype) at a time. An active bind refers to something that is actively consuming and sending data. For example, when you create an IP interface using ifconfig or ipadm, that does an active bind. Another example of an active bind is a daemon used for LLDP. There are also passive binds, like in the case of something trying to capture packets like snoop or tcpdump. That allows something to capture the data without worrying about blocking someone from using that attachment point.

Speeding things up

While the fundamentals of the DLPI are sound, the implementation in STREAMS, particularly for sending data left something to be desired. It greatly complicated the locking and it was hard to get it to perform in the way that was needed for saturating 10 GbE networks with TCP traffic. For all the details on what happened here and a good background, I’ll refer you to Sunay Tripathi’s Blog, where he covers a lot of what changed in Solaris 10 to fix this.

There are two parts to what folks generally end up calling the ‘IP fastpath’. One part of which we leverage for vnd and the other part which is still firmly used by IP. We’ll touch on the first part of this which eliminates the use of sending STREAMS messages. Instead it uses direct callbacks. Today this happens by negotiating with DLPI messages that discover capabilities of devices and then enabling them. The code for the vnd driver does this, as well as the ip driver. Specifically, you first send down a DL_CAPABILITY_REQ message. The response contains a list of capabilities that exist.

If the capability, DL_CAPAB_DLD is returned, then you can enable direct function calls to the DLD and DLS layer. The returned values give you a function pointer, which you can then use to do several things, and ultimately use to request to enable DLD_CAPAB_DIRECT. When you make a call to enable, you specify a function pointer for DLD to call directly when a packet is received. It then will give you a series of functions to use for things like checking flow control and transmitting a packet. These functions allow the system to bypass the issues with STREAMS and directly transmit along packets.

The second part of the ‘IP fastpath’ is something that primarily the IP module uses. In the IP module there is a notion of a neighbor cache entry or nce. This nce describes how to reach another host. When that host is found, the nce asks the lower layers of the stack to generate a layer two header that’s appropriate for this traffic. In the case where you have an Ethernet device, this means that it generates the MAC header including the source and destination mac addresses, ethertype, and vlan tags if there should be one. The IP stack then uses this pre-generated header each time rather than trying to create a new one from scratch for every packet. In addition, the IP module is subscribed to change events that get generated when something like a mac address changes, so that it can regenerate these headers when the administrator makes a change to the system.

New Additions

Finally, it’s worth taking a little bit of time to talk about the new DLPI additions that we added with project bardiche. We needed to solve two problems. Specifically:

To solve the first case, we added a new request called a DL_EXCLUSIVE_REQ. This adds a new mode for the bind state of the dld_str_t. In addition to being active or passive, it can now be exclusive. This can only be requested if no one is actively using the device. If someone is, for example, an IP interface has already been created, then the DL_EXCLUSIVE_REQ will fail. The opposite is true as well, if someone is using the dld_str_t in exclusive mode, then the request to bind to the IP ethertype will also fail. This exclusive request lasts until the consumer closes the dld_str_t.

When a vnd device is created, it makes an explicit request for exclusive access to the device, because it needs to send and receive on all of the different ethertypes. If an IP interface is already active, it doesn’t make sense for a vnd device to be created there. Once the vnd device is destroyed, then anything can use the data link.

Solving our second problem was actually quite simple. The core logic to not loop back packets that were transmitted was already present in the MAC layer. To do that, we created a new promiscuous option that could be specified in the DLPI DL_PROMISCON_REQ called DL_PROMISC_RX_ONLY. Enabling this would pass along the flag MAC_PROMISC_FLAGS_NO_TX_LOOP down to the mac layer which actually does the heavy lifting of duplicating the necessary amount of packets.

Conclusion

This gives a rather rough introduction to the fastpath in the illumos networking stack. The devil, as always, is in the details.

In the next entries, we’ll go over the other new extensions that were added as part of this work: the sdev plugin interface and generalized serialization queues. Finally, we’ll finish with a recap and go over what’s next.

Posted on April 3, 2014 at 8:22 am by rm · Permalink · Leave a comment
In: Miscellaneous

Project Bardiche: Framed I/O

The series so far

If you’re getting started you’ll want to see the previous entries on Project Bardiche:

Background

Framed I/O is a new abstraction that we’re currently experimenting with through Project Bardiche. We call this framed I/O, because the core concept is what we call a frame: a variable amount of discrete data that has a maximum size. In this article, we’ll call data that fits this framed. For example, Ethernet devices work exactly this way. They have a maximum size based on their MTU, but there may very well be less data available than the maximum. There are a few overarching goals that led us down this path:

The primary use case of framed I/O is for vnd and virtual machines. However, we believe that the properties here make it desirable to other portions of the stack which operate in terms of frames. To understand why we’re evaluating this abstraction, it’s worth talking about the other existing abstractions in the system.

read(2) and write(2)

Let’s start with the traditional and most familiar series of I/O interfaces: read(2), write(2), readv(2), and writev(2). These are the standard I/O system calls that most C programmers are familiar with. read(2) and write(2) originated in first edition UNIX. readv(2) and writev(2) supposedly came about during the development of 4.2 BSD. The read and write routines operate on streams of data. The callers and file descriptors have no inherent notion of data being framed and all framing has to be built into consumption of the data. For a lot of use cases, that is the correct abstraction.

The readv(2) and writev(2) interfaces allowed that stream to be vectored. It’s hard to say if these were vectored I/O abstraction in Operating Systems, but it certainly is one of the most popular ones from early systems that’s still around. Where read(2) and write(2) map a stream to a single buffer, these calls map a stream to a series of arbitrarily sized vectors. The act of vectorizing data is not uncommon and can be very useful. Generally, this is done when combining what may be multiple elements into one discrete stream for transferring. For example, if a program maintains one buffer for a protocol header and another buffer is used for the payload, then being able to specify a vector that includes both of these in one call can be quite useful.

When operating with framed data, these interfaces fall a bit short. The problem is that you’ve lost information that the system had regard the framing. It may be that the protocol itself includes the delineations, but there’s no guarantee that that data is correct. For example, if you had a buffer of size 1500, not only would something like read(2) only give you the total number of bytes returned, you wouldn’t be able to get the total number of frames. A return value of 1500 could be one large 1500 byte frame, it could be multiple 300 byte frames or anything in between.

getmsg(2) and putmsg(2)

The next set of APIs that are worth looking at are getmsg(2) and putmsg(2). These APIs are a bit different from the normal read(2) and write(2) APIs, they’re designed around framed messages. These routines use a struct strbuf which has the following members:

    int    maxlen;      /* maximum buffer length */
    int    len;         /* length of data */
    char   *buf;        /* ptr to buffer */

These interfaces allows for the consumer to properly express the maximum size of the frame that they expect and the amount of data that the given frame actually includes. This is very useful for framed data. Unfortunately, this API has some deficiencies. It doesn’t have the ability to break down the data into vectors nor do systems really have a means of working with multiple vectors at a time.

sendmsg(2) and recvmsg(2)

The next set of APIs that I explored and looked at focused on were the recvmsg(2) family, particularly the extensions that were introduced into the Linux kernel via sendmmsg(2) and [recvmmsg(2). The general design of the msghdr structure is good, though it understandably is designed around the socket interface. Unfortunately something like sendmsg(2) is not something that device drivers in most systems get, it currently only works for socket file systems, and a lot of things don't look like sockets. Things like ancillary data and the optional addresses are not as useful and don't have meaning for other styles of messages or if they do, they may not fit the abstraction that's been defined there.

Framed I/O

Based on our evaluations with the above APIs, a few of us chatted around Joyent's San Francisco office and tried to come up with something that might have the properties we felt made more sense for something like KVM networking. To help distinguish it from traditional Socket semantics or STREAMS semantics, we named it after the basic building block of the frame. The general structure is called a frameio_t which itself has a series of vector structures called a framevec_t. The structures roughly look like:

typedef struct framevec {
    void    *fv_buf;        /* Buffer with data */
    size_t  fv_buflen;      /* Size of the buffer */
    size_t  fv_actlen;      /* Amount of buffer consumed, ignore on error */
} framevec_t;

typedef struct frameio {
    uint_t  fio_version;    /* Should always be FRAMEIO_CURRENT_VERSION */
    uint_t  fio_nvpf;       /* How many vectors make up one frame */
    uint_t  fio_nvecs;      /* The total number of vectors */
    framevec_t fio_vecs[];  /* C99 VLA */
} frameio_t;

The idea here is that, much like a struct msgbuf, each vector component has a notion of what it's maximum size is and then the actual size of data consumed. These vectors can then be constructed into series of frames in multiple ways through the fio_nvpf and fio_nvecs members. The fio_nvecs field describes the total number of vectors and the fio_nvpf describes how many vectors are in a frame. You might think of the fio_nvpf member as basically describing how many iovec structures make up a single frame.

Consider that you have four vectors to play with, you might want to rig it up in one of several ways. You might want to map each message to a signle vector, meaning that you could read four messsages at once. You might want the opposite and map a single message to all four vectors. In that case you'd only ever read one message at a time, broken into four components. You could also break it down such that you always broke down a message into two vectors, that means that you'd be able to read two messages at a time. The following ASCII art might help.

1:1 Vector to Frame mapping

 +-------+  +-------+  +-------+  +-------+
 | msg 0 |  | msg 1 |  | msg 2 |  | msg 3 |
 +-------+  +-------+  +-------+  +-------+
    ||         ||         ||         ||
 +-------+  +-------+  +-------+  +-------+
 | vec 0 |  | vec 1 |  | vec 2 |  | vec 3 |
 +-------+  +-------+  +-------+  +-------+

4:1 Vector to Frame mapping

 +----------------------------------------+
 |                  msg 0                 |
 +----------------------------------------+
    ||         ||         ||         ||
 +-------+  +-------+  +-------+  +-------+
 | vec 0 |  | vec 1 |  | vec 2 |  | vec 3 |
 +-------+  +-------+  +-------+  +-------+

2:1 Vector to Frame Mapping

 +------------------+  +------------------+
 |       msg 0      |  |       msg 1      |
 +------------------+  +------------------+
    ||         ||         ||         ||
 +-------+  +-------+  +-------+  +-------+
 | vec 0 |  | vec 1 |  | vec 2 |  | vec 3 |
 +-------+  +-------+  +-------+  +-------+

Currently the maximum number of vectors allowed in a given call is limited to 32. As long as the total number evenly divides the number of frames per vector, than any value is alright.

By combining these two different directions, we believe that this'll be a useful abstraction and suitable for other parts of the system that operate on framed data, for example a USB stack. Another thing that this design lets us do is that by not constraining the content of the vectors, it would be possible to replicate something like the struct msghdr where the protocol header data was actually in the first vector.

Framed I/O now and in the Future

Today we've started plumbing this through in QEMU to account for its network device backend APIs that allow one to operate on traditional iovec. However, there's a lot more that can be done with this. For example, one of the things that is on our minds is writing a vhost-net style driver for illumos that can easily map data between the framed I/O representation and the virtio driver. With this, it'd even be possible to do something that's mostly zero-copy as well. Alternatively, we may also explore just redoing a lot of QEMU's internal networking paths to make it more friendly for sending and receiving multiple packets at once. That should certainly help with the overhead today of networking I/O in virtual machines.

We think that this might fit in other parts of the system as well, for example, it may make sense to be used as part of the illumos USB3 stack's design as the unit that we send data in. Whether it makes sense as anything more than just for the vnd device and this style of I/O time will tell.

Today vnd devices are exposed in libvnd through the vnd_frameio_read(3VND) and vnd_frameio_write(3VND) interfaces. So these can also be used for someone who's trying to develop their own services using vnd, for example, user land switches, firewalls, etc.

Next in the Bardiche Series

Next in the bardiche series, we'll be delving into some of the additional kernel subsystems and new DLPI abstractions that were created. Following those, we'll end with a recap entry on bardiche as a whole and what may come next.

Posted on March 25, 2014 at 8:57 am by rm · Permalink · Leave a comment
In: Miscellaneous

Project Bardiche: vnd Architecture

My previous entry introduced Project Bardiche, a project which revamps how we do networking for KVM guests. This entry focuses on the design and architecture of the vnd driver and how it fits into the broader networking stack.

The illumos networking stack

The illumos networking stack is broken into several discrete pieces which is summarized in the following diagram:

             +---------+----------+----------+
             | libdlpi |  libvnd  | libsocket|
             +---------+----------+----------+
             |         ·          ·    VFS   |
             |   VFS   ·    VFS   +----------+
             |         ·          |  sockfs  |
             +---------+----------+----------+
             |         |    VND   |    IP    |
             |         +----------+----------+
             |            DLD/DLS            |
             +-------------------------------+
             |              MAC              |
             +-------------------------------+
             |             GLDv3             |
             +-------------------------------+

At the top of the diagram are the common interfaces to user land. The first and most familiar is libsocket. That contains all of the common interfaces that C programmers are used to seeing: connect(3SOCKET), accept(3SOCKET), etc. On the other hand, there’s libdlpi, which provides an alternate means for interacting with networking devices. It is often used by software like DHCP servers and for LLDP. libvnd is new and a part of project bardiche. For now, we’ll stick to describing the other two paths.

Next, operations transit through the virtual file system (VFS) to reach their next destination. For most folks, that’ll be the illumos file system sockfs. The sockfs file system provides a means of translating between the gory details of TCP, UDP, IP, and friends, and the actual sockets that they rely on. The next step for such sockets is what folks traditionally think of as the TCP/IP stack. This encompasses everything related to the actual protocol processing. For example, connecting a socket and going through the TCP dance of the SYN and SYN/ACK is all handled by the logic in the TCP/IP stack. In illumos, both TCP and IP are implemented in the same kernel module called IP.

The next layer is comprised of two kernel modules which work together called DLD and DLS. DLD is the data-link driver and DLS is the data-link services module. The two modules work together. Every data link in illumos, whether it’s a virtual nic or physical nic, is modeled as a dld device. When you open something like /dev/net/igb0, that’s an instance of a DLD device. These devices provide an implementation of all of the DLPI (Data-link Provider Interface) STREAMS messages and are used to negotiate the fast path. We’ll go into more detail about that in a future entry.

Everything transits out of DLD and DLS and enters the MAC layer. The MAC layer handles taking care of interfacing with the actual device drivers, programming unicast and multicast addresses into devices, controlling whether or not the devices are in promiscuous mode, etc. The final layer is the Generic Lan Device version three (GLDv3) Device Driver. GLDv3 is a standard interface for networking device drivers and represents a set of entry points that the Operating System expects to use with them.

vnd devices

A vnd device is created on top of a data link similar to how an IP interface is created on top of a data link. Once a vnd device has been created, it can be used to read and write layer two frames. In addition, a vnd device can optionally be linked into the file system name space allowing others to open the device.

Similar to /dev/net, vnd devices show up under /dev/vnd. A control node is always created at /dev/vnd/ctl. This control node is referred to as a self-cloning device. That means that any time the device is opened, a new instance of the device is created. Once the control node has been opened, it is associated with a data link and then it is bound into the file system name space with some name that usually is identical to the name of the data link. After the device has been bound, it then shows up in /dev/vnd. If a vnd device was named net0 then it would show up as /dev/vnd/net0. Just as /dev/net displays all of the data links in th various zones under /dev/net/zone, the same is true for vnd. The vnd devices in any zone are all located under /dev/vnd/zone and follow the pattern /dev/vnd/zone/%zonename/%vnddevice. These devices are never directly manipulated. Instead, they are used by libvnd and vndadm.

Once a vnd device has been created and bound into the name space, it will persist until it is removed with either vndadm or libvnd or the zone it is present in is halted. The removal of vnd devices from the name space is similar to calling unlink(2) on a file. If any process has the vnd device open after it is has been removed from the name space, it will persist until all open handles have been closed.

If a data link already has an IP interface or is being actively used for any other purpose, a vnd device cannot be created on top of it, and vice versa. Because vnd devices operate at a layer two, if various folks are already consuming layer three, it doesn’t make sense to create a vnd device on top of it. The opposite also holds.

The command vndadm was written to manipulate vnd devices. It’s worth stepping through some basic examples of using the command. Even more examples can be found in its manual page. With that, let’s get started and create a vnic and then a device. Substitute your physical link for anything you prefer.

# dladm create-vnic -l e1000g0 vnd0
# vndadm create vnd0
# ls /dev/vnd
ctl   vnd0  zone

With that, we’ve created a device. Next, we can use vndadm to list devices as well as get and set properties.

# vndadm list
NAME             DATALINK         ZONENAME
vnd0             vnd0             global
# vndadm get vnd0
LINK          PROPERTY         PERM  VALUE
vnd0          rxbuf            rw    65536
vnd0          txbuf            rw    65536
vnd0          maxsize          r-    4194304
vnd0          mintu            r-    0
vnd0          maxtu            r-    1518
# vndadm set txbuf=2M
# vndadm get vnd0 txbuf
LINK          PROPERTY         PERM  VALUE
vnd0          txbuf            rw    2097152

You’ll note that there are two properties that we can set, rxbuf and txbuf. These are the sizes of buffers that an instance of a vnd device maintains. As frames come in, they are put into the receive buffer where they sit until they are read by someone, usually a KVM guest. If a frame would come in that would exceed the size of that buffer, then it will be dropped instead. The transmit buffer controls the total amount of outstanding data that can exist at any given time in the vnd subsystem. The vnd device has to keep track of this to deal with cases like flow control.

Finally, we can go ahead and remove the device via:

# vndadm destroy vnd0

While not shown here, all of these commands can operate on device that are in another zone, if the user is in the global zone. To get statistics about device throughput and packet drops, you can use the command vndstat. Here’s a brief example:

$ vndstat 1 5
 name |   rx B/s |   tx B/s | drops txfc | zone
 net0 | 1.45MB/s | 14.1KB/s |     0    0 | 1b7155a4-aef9-e7f0-d33c-9705e4b8b525
 net0 | 3.50MB/s | 19.5KB/s |     0    0 | 1b7155a4-aef9-e7f0-d33c-9705e4b8b525
 net0 | 2.83MB/s | 30.8KB/s |     0    0 | 1b7155a4-aef9-e7f0-d33c-9705e4b8b525
 net0 | 3.08MB/s | 30.6KB/s |     0    0 | 1b7155a4-aef9-e7f0-d33c-9705e4b8b525
 net0 | 3.21MB/s | 30.6KB/s |     0    0 | 1b7155a4-aef9-e7f0-d33c-9705e4b8b525

The drops column sums up the total number of drops while the txfc column shows the number of times that the device has been flow controlled during that period.

Programmatic Use

So far, I’ve demonstrated the use of the user commands. For most applications, you’ll want to use the fully featured C library libvnd. The introductory manual is where you’ll want to get started for all information in using it. It will point you out to all of the rest of the functions which can all be found in the manual section 3VND. Please keep in mind, that until the library makes its way up into illumos, portions of the API may still end up changing and isn’t considered stable yet.

Peeking under the hood

So far we’ve talked about how you can use these devices, now let’s go under the hood and talk about how this is constructed. For all of the full gory details, you should turn to the vnd big theory statement. There are multiple components that make up the general architecture of the vnd sub-system, though only character devices are shown. The following bit of ascii art from the big theory statement describes the general architecture:

+----------------+     +-----------------+
| global         |     | global          |
| device list    |     | netstack list   |
| vnd_dev_list   |     | vnd_nsd_list    |
+----------------+     +-----------------+
    |                    |
    |                    v
    |    +-------------------+      +-------------------+
    |    | per-netstack data | ---> | per-netstack data | --> ...
    |    | vnd_pnsd_t        |      | vnd_pnsd_t        |
    |    |                   |      +-------------------+
    |    |                   |
    |    | nestackid_t    ---+----> Netstack ID
    |    | vnd_pnsd_flags_t -+----> Status flags
    |    | zoneid_t       ---+----> Zone ID for this netstack
    |    | hook_family_t  ---+----> VND IPv4 Hooks
    |    | hook_family_t  ---+----> VND IPv6 Hooks
    |    | list_t ----+      |
    |    +------------+------+
    |                 |
    |                 v
    |           +------------------+       +------------------+
    |           | character device |  ---> | character device | -> ...
    +---------->| vnd_dev_t        |       | vnd_dev_t        |
                |                  |       +------------------+
                |                  |
                | minor_t       ---+--> device minor number
                | ldi_handle_t  ---+--> handle to /dev/net/%datalink
                | vnd_dev_flags_t -+--> device flags, non blocking, etc.
                | char[]        ---+--> name if linked
                | vnd_str_t * -+   |
                +--------------+---+
                               |
                               v
        +-------------------------+
        | STREAMS device          |
        | vnd_str_t               |
        |                         |
        | vnd_str_state_t      ---+---> State machine state
        | gsqueue_t *          ---+---> mblk_t Serialization queue
        | vnd_str_stat_t       ---+---> per-device kstats
        | vnd_str_capab_t      ---+----------------------------+
        | vnd_data_queue_t ---+   |                            |
        | vnd_data_queue_t -+ |   |                            v
        +-------------------+-+---+                  +---------------------+
                            | |                      | Stream capabilities |
                            | |                      | vnd_str_capab_t     |
                            | |                      |                     |
                            | |    supported caps <--+-- vnd_capab_flags_t |
                            | |    dld cap handle <--+-- void *            |
                            | |    direct tx func <--+-- vnd_dld_tx_t      |
                            | |                      +---------------------+
                            | |
           +----------------+ +-------------+
           |                                |
           v                                v
+-------------------+                  +-------------------+
| Read data queue   |                  | Write data queue  |
| vnd_data_queue_t  |                  | vnd_data_queue_t  |
|                   |                  |                   |
| size_t        ----+--> Current size  | size_t        ----+--> Current size
| size_t        ----+--> Max size      | size_t        ----+--> Max size
| mblk_t *      ----+--> Queue head    | mblk_t *      ----+--> Queue head
| mblk_t *      ----+--> Queue tail    | mblk_t *      ----+--> Queue tail
+-------------------+                  +-------------------+

At a high level there are three different core components. There is a per-netstack data structure, there is a character device and there is a STREAMS device.

A netstack, or networking stack, is a concept in illumos that contains an independent set of networking information. This includes TCP/IP state, routing tables, tunables, etc. Every zone in SmartOS has its own netstack which allows zones to more fully control and interface with networking. In addition, the system has a series of IP hooks which are used by things like ipfilter and ipd to manipulate packets. When the vnd kernel module is first loaded, it registers with the netstack sub-system which ensures that allows the vnd kernel module to create its per-netstack data. In addition to hooking, the per-netstack data is used to make sure that when a zone is halted that all of the associated vnd devices are torn down.

The character device is the interface between consumers and the system. The vnd module is actually a self-cloning device. Whenever a library handle is created it first opens the control node which is /dev/vnd/ctl. The act of opening that creates a clone of the device with a new minor number. When an existing vnd device is opened, then no cloning takes place, it opens one of the existing character devices.

The major magic happens when a vnd character device is asked to associate with a data link. This happens through an ioctl that the library wraps up and takes care of. When the device is associated, the kernel itself does what we call a layered open – it opens and holds another character or block device. In this case the vnd module does a layered open of the data link. However, the devices that back data links are still STREAMS devices that speak DLPI. To take care of dealing with all of the DLPI messages and set up the normal fast path, we use the third core component: the vnd STREAMS device.

The vnd STREAMS device is fairly special, it cannot be used outside of the kernel and is an implementation detail of the vnd driver. After doing the layered open, the vnd STREAMS device is pushed onto the stream head and it begins to exchange DLPI messages to set up and configure the data link. Once it has successfully walked through its state machine, the device is full and ready to go. As part of doing that, it asks for exclusive access to the device, enables us to receive all the packets that are originally destined for the device, and enables direct function calls for this through what’s referred to commonly as the fastpath. Once that’s set up, the character device and STREAMS device wire up with one another. Once that’s all been finished successfully, the character device can be fully initialized.

At this point in time, the device can be fully used for reading and writing packets. It can optionally be bound into the file system name space. That binding is facilitated by the sdev file system and its new plugin interface. We’ll go into more detail about that in a future entry.

The STREAMS device contains a lot of the meat for dealing with data. It contains the data queues and it controls all the interactions with DLD/DLS and the fastpath. In addition, it also knows about its gsqueue (generic serialization queue). The gsqueue is used to ensure that we properly handle the order of transmitted packets, especially when subject to flow control.

The following two diagrams (from the big theory statement) describe the path that data takes when received and when transmitted.

Receive path

                                 |
                                 * . . . packets from gld
                                 |
                                 v
                          +-------------+
                          |     mac     |
                          +-------------+
                                 |
                                 v
                          +-------------+
                          |     dld     |
                          +-------------+
                                 |
                                 * . . . dld direct callback
                                 |
                                 v
                         +---------------+
                         | vnd_mac_input |
                         +---------------+
                                 |
                                 v
  +---------+             +-------------+
  | dropped |<--*---------|  vnd_hooks  |
  |   by    |   .         +-------------+
  |  hooks  |   . drop probe     |
  +---------+     kstat bump     * . . . Do we have free
                                 |         buffer space?
                                 |
                           no .  |      . yes
                              .  +      .
                          +---*--+------*-------+
                          |                     |
                          * . . drop probe      * . . recv probe
                          |     kstat bump      |     kstat bump
                          v                     |
                       +---------+              * . . fire pollin
                       | freemsg |              v
                       +---------+   +-----------------------+
                                     | vnd_str_t`vns_dq_read |
                                     +-----------------------+
                                              ^ ^
                              +----------+    | |     +---------+
                              | read(9E) |-->-+ +--<--| frameio |
                              +----------+            +---------+

Transmit path

  +-----------+   +--------------+       +-------------------------+   +------+
  | write(9E) |-->| Space in the |--*--->| gsqueue_enter_one()     |-->| Done |
  | frameio   |   | write queue? |  .    | +->vnd_squeue_tx_append |   +------+
  +-----------+   +--------------+  .    +-------------------------+
                          |   ^     .
                          |   |     . reserve space           from gsqueue
                          |   |                                   |
             queue  . . . *   |       space                       v
              full        |   * . . . avail          +------------------------+
                          v   |                      | vnd_squeue_tx_append() |
  +--------+          +------------+                 +------------------------+
  | EAGAIN |<--*------| Non-block? |<-+                           |
  +--------+   .      +------------+  |                           v
               . yes             v    |     wait          +--------------+
                           no . .*    * . . for           | append chain |
                                 +----+     space         | to outgoing  |
                                                          |  mblk chain  |
    from gsqueue                                          +--------------+
        |                                                        |
        |      +-------------------------------------------------+
        |      |
        |      |                            yes . . .
        v      v                                    .
   +-----------------------+    +--------------+    .     +------+
   | vnd_squeue_tx_drain() |--->| mac blocked? |----*---->| Done |
   +-----------------------+    +--------------+          +------+
                                        |                     |
      +---------------------------------|---------------------+
      |                                 |           tx        |
      |                          no . . *           queue . . *
      | flow controlled .               |           empty     * . fire pollout
      |                 .               v                     |   if mblk_t's
    +-------------+     .      +---------------------+        |   sent
    | set blocked |<----*------| vnd_squeue_tx_one() |--------^-------+
    | flags       |            +---------------------+                |
    +-------------+    More data       |    |      |      More data   |
                       and limit       ^    v      * . .  and limit   ^
                       not reached . . *    |      |      reached     |
                                       +----+      |                  |
                                                   v                  |
    +----------+          +-------------+    +---------------------------+
    | mac flow |--------->| remove mac  |--->| gsqueue_enter_one() with  |
    | control  |          | block flags |    | vnd_squeue_tx_drain() and |
    | callback |          +-------------+    | GSQUEUE_FILL flag, iff    |
    +----------+                             | not already scheduled     |
                                             +---------------------------+

Wrapping up

This entry introduces the tooling around vnd and provides a high level overview of the different components that make up the vnd module. In the next entry in the series on baridche, we’ll cover the new framed I/O abstraction. Entries following that will cover the new DLPI extensions, the sdev plugin interface, generalized squeues, and finally the road ahead.

Posted on March 24, 2014 at 8:31 am by rm · Permalink · Leave a comment
In: Bardiche · Tagged with: , , ,

Project Bardiche: Introduction

I just recently landed Project Bardiche into SmartOS. The goal of Bardiche has been to create a more streamlined data path for layer two networking in illumos. While the primary motivator for this was for KVM guests, it’s opened up a lot of room for more than just virtual machines. This bulk of this project is comprised of changes to illumos-joyent; however, there were some minor changes made to smartos-live, illumos-kvm, and illumos-kvm-cmd.

Project Highlights

Before we delve into a lot more of the specifics in the implementation, let’s take a high level view of what this project has brought to the system. Several of these topics will have their own follow up blog entries.

There’s quite a bit there. The rest of this entry will go into detail on the motivation for this work and a bit more on the new /dev/net/zone, libdlpi, and snoop features. Subsequent entries will cover the the new vnd architecture and the new DLPI primitives, the new gsqueue interface, shine a bit more light on what gets referred to as the fastpath, and cover the new sdev plugin architecture.

Motivation

Project bardiche started from someone asking what would it take to allow a hypervisor-based firewall to be able to filter packets that were being sent from a KVM guest. We wanted to focus on allowing the hypervisor to provide the firewall because of the following challenges associated with managing a firewall running in the guest.

While it’s true that practically all the guests that you would run under hardware virtualization have their own firewall software, they’re rarely the same. If we wanted to leverage the firewall built into the guest, we’d need to build an agent that lived in each guest. Not only does that mean that we’d have to write one of these for every type of guest we wanted to manage, given that customers are generally the super-user in their virtual machine (VM), they’d be able to simply kill the agent or change these rules, defeating the API.

While dealing with this, there were several other deficiencies in how networking worked for KVM guests today based on how QEMU, the program that actually runs the VM, interacted with the host networking. For each Ethernet device that was in the guest, there was a corresponding virtual NIC in the host. The two were joined with the vnic back end in QEMU which originally used libdlpi to bind them. While this worked, there were some problems with it.

Because we had to put the device in promiscuous mode, there was no way to tell it not to send back traffic that came from ourselves. In addition to just being a waste of cycles, this causes duplicate address detection, often performed with IPv6, to fail for many systems.

In addition, the dlpi interfaces had no means of reading or writing multiple packets at a time. A lot of these issues extend from the history of the illumos networking stack. When it was first implemented, it was done using STREAMS. Over time, that has changed. In Solaris 10 the entire networking stack was revamped with a project called Fire Engine. That project, among many others, transitioned the stack from a message passing interface to one that used a series of direct calls and serialization queues (squeues). Unfortunately, the means by which we were using libdlpi, left us still using STREAMS.

While exploring the different options and means to interface with the existing firewall, we eventually reached the point where we realized that we needed to go out and create a new interface that solved this, and the related problems that we had, as well as, lay the foundation for a lot of work that we’d still like to do.

First Stop: Observability Improvements

When I first started this project, I knew that I was going to have to spend a lot of time debugging. As such, I knew that I need to solve one of the more frustrating aspects of working with KVM networking: the ability to snoop and capture traffic. At Joyent, we always run a KVM instance inside of zone. This gives us all the benefits of zones: the associated security and resource controls.

However, before this project, data links that belonged to zones were not accessible from the global zone. Because of the design of the KVM branded zone, the only process running is QEMU and you cannot log in, which makes it very hard to pass the data link to snoop or tcpdump. This set up does not make it impossible to debug. One can use DTrace or use snoop on a device in the global zone; however, both of those end up requiring a bit more work or filtering.

The solution to this is to allow the global zone to see the data links for all devices across all zones under /dev/net and then enhance the associated libraries and commands to support accessing the new devices. If you’re in the global zone, there is now a new directory called /dev/net/zone. Don’t worry, this new directory can’t break you, as all data links in the system need to end with a number. On my development virtual machine which has a single zone with a single vnic named net0, you’d see the following:

[root@00-0c-29-37-80-28 ~]# find /dev/net | sort
/dev/net
/dev/net/e1000g0
/dev/net/vmwarebr0
/dev/net/zone
/dev/net/zone/79809c3b-6c21-4eee-ba85-b524bcecfdb8
/dev/net/zone/79809c3b-6c21-4eee-ba85-b524bcecfdb8/net0
/dev/net/zone/global
/dev/net/zone/global/e1000g0
/dev/net/zone/global/vmwarebr0

Just as you always have in SmartOS, you’ll still see the data links for your zone at the top level in /dev/net, eg. /dev/net/e1000g0 and /dev/net/vmwarebr0. Next, each of the zones on the system, in this case the global zone, and the non-global zone named 79809c3b-6c21-4eee-ba85-b524bcecfdb8, show up in /dev/net/zone. Inside of each of those directories are the data links that live in that zone.

The next part of this was exposing this functionality in libdlpi and then using that in snoop. For the moment, I added a private interface called dlpi_open_zone. It’s similar to dlpi_open except that it takes an extra argument for the zone name. Once this change gets up to illumos it’ll then become a public interface that you can and should use. You can view the manual page online here or if you’re on a newer SmartOS box you can run man dlpi_open_zone to see the documentation.

The use of this made its way into snoop in the form of a new option: -z zonename. Specifying a zone with -z will cause snoop to use dlpi_open_zone which will try to open the data link specified via its -d option from the zone. So if we wanted to watch all the icmp traffic over the data link that the KVM guest used we could run:

# snoop -z 79809c3b-6c21-4eee-ba85-b524bcecfdb8 -d net0 icmp

With this, it should now be easier, as an administrator of multiple zones, to observe what’s going on across multiple zones without having to log into them.

Thanks

There are numerous people whom helped this project along the way. The entire Joyent engineering team helped from the early days of bouncing ideas about APIs and interfaces all the way through to the final pieces of review. Dan McDonald, Sebastien Roy, and Richard Lowe, all helped review the various write ups and helped deal with various bits of arcana in the networking stack. Finally, the broader SmartOS community helped go through and provide additional alpha and beta testing.

What’s Next

In the following entries, we’ll take a tour of the new sub-systems ranging from the vnd architecture and framed I/O abstraction through the sdev interfaces.

Posted on March 20, 2014 at 4:01 pm by rm · Permalink · Comments Closed
In: Bardiche · Tagged with: , , , ,

Userland CTF in DTrace

We at Joyent use DTrace for understanding and debugging userland applications just as often as we do for the kernel. That is part of the reason why we’ve worked on things like flamegraphs, the Node.js ustack helper, and the integration of libusdt in node module’s like restify and bunyan.

I’ve just put back some work that makes observing userland with DTrace a lot simpler and much more powerful. Before we meet the devil in the details, let’s start with an example:

$ cat startd.d
pid$target::start_instance:entry,
pid$target::stop_instance:entry
{
        printf("%s: %s\n", probefunc == "start_instance" ?
            "starting" : "stopping", stringof(args[1]->ri_i.i_fmri));
}
$ dtrace -qs startd.d -p $(pgrep -o startd)
...
stopping: svc:/system/intrd:default
stopping: svc:/system/cron:default
start_instance:entry starting: svc:/system/cron:default
...

If you’re familiar with DTrace you might realize that this doesn’t really look like what you’re expecting! Hey Robert, won’t that script blow up without the copyin? Also where did the args[] come from with the pid provider?!

The answer to this minor mystery is that DTrace can now leverage type information in userland programs. Not only does the compiler know the size and layout of types, it’ll also take care of adding calls to copyin for you so you can dereference members without fear. To explain how we’ve managed all of this, we need to go into a bit of back story.

The Compiler is the Enemy

Since the beginning of programming, we’ve needed to be able to debug the programs that we’ve written. A large chunk of time has been spent on tooling to be able to understand and root cause these bugs whether working on a live system or doing a post-mortem analysis on something like a core dump.

Unfortunately, the compiler is in many ways our enemy. Its goal is to take whatever well commented and understandable code we might have and transform it not only into something that a processor can understand, but often times transforming it through optimization passes into something that no longer resembles what we originally wrote.

This problem isn’t limited to languages like C and C++. In fact, many of the same problems apply when you use any compiler, be it your current compile to JavaScript language of the day (emscripten and coffeescript) or something like lex and yacc.

Fortunately, both the compiler and the linker are just software. Shortly after we first hit this problem, they were modified to produce debugging information that could be encoded into the binaries they produced. Folks even were able to encode this kind of information in the original ‘a.out’ executable format that came around in first edition UNIX.

One of the first popular formats that was used was called stabs. It was used on many operating systems for many years and you can still convince compilers like gcc, clang, and suncc to generate it. Since then, DWARF has become the most popular and commonly used format. The initial origin of DWARF came from Bell Labs, but it was rather unpopular because the debugging data that it created was just too large. Since then DWARF has become more standardized and more compact than the first version of DWARF. However, it is a rather complicated format.

With all of these formats there is a trade-off between expressibility and size. If the debugging information takes too much space, then people stop including it. Most available OS and package distributions do not incorporate debugging information. If you’re lucky, they separate that information into different packages. This means that when you’re debugging a problem in production you very often don’t have the very information you need. Even more frustrating, when this information is in a separate file and you’re trying to do post-mortem analysis, then you need to track that down and make sure that you have the right version of the debug information that corresponds to what you were using in production.

This situation is unsatisfying, but – we have other options! Sun developed CTF in Solaris 9 with the intent of using it with mdb and eventually DTrace. In illumos, we put CTF data in every kernel module, library, and a majority of applications. We don’t store all the information that you might get in, say, DWARF, but we store what we’ve found over the years to be the most useful information.

CTF data includes the following pieces of information:

o The definitions of all types and structures
o The arguments and types of each function
o The types of function return values
o The types of global variables

All of the CTF data for a given library, binary, or kernel module is found inside of what we call a CTF Container. The CTF container is found as its own section in an ELF object. A simple way to see if something in question has CTF is to run elfdump(1). Here’s an example:

$ elfdump /lib/libc.so | grep SUNW_ctf
Section Header[37]:  sh_name: .SUNW_ctf

If a library or program does not have CTF data then the section won’t show up in the list and the grep should turn up empty.

CTF and DTrace

If you’ve ever wondered how it is that DTrace knows types when you use various providers, like args[] in fbt, the answer to that is CTF. When you run DTrace, it loads relevant CTF containers for the kernel. In fact, even the basic types that D provides, such as an int or types that you define in a D script, end up in a CTF container. Consider the following dtrace invocation:

# dtrace -l -v -n 'fbt::ip_input:entry'
   ID   PROVIDER            MODULE                          FUNCTION
NAME
37546        fbt                ip                          ip_input
entry

        Probe Description Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: Unknown

        Argument Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: ISA

        Argument Types
                args[0]: struct ill_s *
                args[1]: ill_rx_ring_t *
                args[2]: mblk_t *
                args[3]: struct mac_header_info_s *

The D compiler used its CTF data for the ip module to determine the arguments and their types. We can then run something like:

# dtrace -qn 'fbt::ip_input:entry{ print(*args[0]); exit(0) }'
...
struct ill_s {
    pfillinput_t ill_inputfn = ip`ill_input_short_v4
    ill_if_t *ill_ifptr = 0xffffff0148a27ab8
    queue_t *ill_rq = 0xffffff014b65ba60
    queue_t *ill_wq = 0xffffff014b65bb58
    int ill_error = 0
    ipif_t *ill_ipif = 0xffffff014b6fc460
    uint_t ill_ipif_up_count = 0x1
    uint_t ill_max_frag = 0x5dc
    uint_t ill_current_frag = 0x5dc
    uint_t ill_mtu = 0x5dc
    uint_t ill_mc_mtu = 0x5dc
    uint_t ill_metric = 0
    char *ill_name = 0xffffff014bc642c8
...
}

DTrace uses the CTF data for the struct ill_s to interpret all of the data and correlate it with the appropriate members.

Bringing it to Userland

While DTrace happily consumes all of the CTF data for the various kernel modules, up to now it simply ignored all of the CTF data in userland applications. With my changes, DTrace will now consume that CTF data for referenced processes. Let’s go back to the example that we opened this blog post with. If we list those probes verbosely we see:

# dtrace -l -v -n
# 'pid$target::stop_instance:entry,pid$target::start_instance:entry' -p
# $(pgrep -o startd)
   ID   PROVIDER            MODULE                          FUNCTION
NAME
62420       pid8        svc.startd                     stop_instance
entry

        Probe Description Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: Unknown

        Argument Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: Unknown

        Argument Types
                args[0]: userland scf_handle_t *
                args[1]: userland restarter_inst_t *
                args[2]: userland stop_cause_t

62419       pid8        svc.startd                    start_instance
entry

        Probe Description Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: Unknown

        Argument Attributes
                Identifier Names: Private
                Data Semantics:   Private
                Dependency Class: Unknown

        Argument Types
                args[0]: userland scf_handle_t *
                args[1]: userland restarter_inst_t *
                args[2]: userland int32_t

Before this change, the Argument Types section would be empty. Because svc.startd has CTF data, DTrace was able to figure out the types of startd’s functions. Without these changes, you’d have to manually define the types of a scf_handle_t and a restarter_inst_t and cast the raw arguments to the correct type. If you ended up with a structure that has a lot of nested structures, defining all of them in D can quickly become turtles all the way down.

Look Ma, no copyin!

DTrace runs in a special context in the kernel, and often times DTrace requires you to think about what’s going on. Just as the kernel can’t access arbitrary user memory without copying it in, neither can DTrace. Consider the following classic one liner:

dtrace -n 'syscall::open:entry{ trace(copyinstr(arg0)); }'

You’ll note that we have to use copyinstr. That tells DTrace that we need to copy in the string from userland into the kernel in order to do something with it, whether that be an aggregation, printing it out, or saving it for some later action. This copyin isn’t limited to just strings. If you wanted to dereference some member of a structure, you’d have to either copy in the full structure, or manually determine the offset location of the member you care about.

At the previous illumos hackathon, Adam Leventhal had the idea of introducing a keyword into D, the DTrace language, that would tell the D compiler that a type is from userland. The D compiler would then take care of copying in the data automatically. Together we built a working prototype, with the keyword userland.

While certainly useful on its own, it really shines when combined with CTF data, as in the pid provider. The pid provider automatically applies the userland keyword to all of the types that are found in args[]. This allowed us to skip the copyin of intermediate structures and write a simple expression. eg. in our initial example we are able to do something that looks like a normal dereference in C: args[1]->ri_i.i_fmri. Before this change, you would have had to do three copyins: one for args[1], one for ri_i, and a final one for the string i_fmri.

As an example of the kinds of scripts that motivated this, here’s a portion of a D script that I used to help debug part of an issue inside of QEMU and KVM:

$ cat event.d
...
/* After its mfence */
pid$target::vring_notify:entry
/self->trace/
{
    self->arg = arg1;
    this->data = (char *)(copyin(self->arg + 0x28, 8));
    self->sign = *(uint16_t *)(this->data+0x2);
}

...

pid$target::vring_notify:return
/self->trace && self->arg/
{
    this->data =  (char *)(copyin(self->arg + 0x28, 8));
    printf("%d notify signal index: 0x%04x notify? %d\n", timestamp,
        *(uint16_t *)(this->data + 0x2), arg1);
}
...

There are many parts of this script where I’ve had to manually encode structure offsets and structure sizes. In the larger script, I had to play games with looking at registers and the pid provider’s ability to instrument arbitrary assembly instructions. I for one am glad that I’ll have to write a lot less of these.

When you have no CTF

While no binary should be left behind, not everything has CTF data today. But the userland keyword can still be incredibly useful even without it. Whenever you’re making a cast, you can note that the type is a userland type with the userland keyword, and the D compiler will do all the heavy lifting from there.

Here’s an example from a program that has a traditional linked list, but doesn’t have any CTF data:

struct foo;

typedef struct foo {
        struct foo *foo_next;
        char *foo_value;
} foo_t;

pid$target::print_head:entry
{
        this->p = (userland foo_t *)arg0;
        trace(stringof(this->p->foo_next->foo_next->foo_value));
}

The nice thing with the userland keyword is that you don’t have to do any copyin or worry about figuring out structure sizes. The goal with all of this is to make it simpler and more intuitive to write D scripts.

Referring to types

As a part of all this, you can refer to types in arbitrary processes that are running on the system, as well as the target. The syntax is designed to be flexible enough to allow you to specify not just the pid, but the link map, library, and type name, but you can also just specify the pid and type that you want. While you can’t refer to macros inside these definitions, you can use the shorthand `pid“ to refer to the value of $TARGET.

For example, say you wanted to refer to a glob_t, which on illumos is defined via a typdef struct glob_t { ... } glob_t, there are a lot of different ways that you can do that. The following are all equivalent:

dtrace -n 'BEGIN{ trace((pid`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((pid`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((pid`LM0`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((pid8`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((pid8`libc.so.1`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((pid8`LM0`libc.so.1`glob_t *)0); }'

dtrace -n 'BEGIN{ trace((struct pid`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct pid`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct pid`LM0`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct pid8`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((struct pid8`libc.so.1`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((struct pid8`LM0`libc.so.1`glob_t *)0); }'

All of these would also work with the userland keyword. The userland keyword interacts with structs a bit differently than one might expect, so let’s show all of our above examples with the userland keyword:

dtrace -n 'BEGIN{ trace((userland pid`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((userland pid`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((userland pid`LM0`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((userland pid8`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((userland pid8`libc.so.1`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((userland pid8`LM0`libc.so.1`glob_t *)0); }'

dtrace -n 'BEGIN{ trace((struct userland pid`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct userland pid`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct userland pid`LM0`libc.so.1`glob_t *)0); }' -p 8
dtrace -n 'BEGIN{ trace((struct userland pid8`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((struct userland pid8`libc.so.1`glob_t *)0); }'
dtrace -n 'BEGIN{ trace((struct userland pid8`LM0`libc.so.1`glob_t *)0); }'

What’s next?

From here you can get started with userland ctf and the userland keyword in DTrace in current releases of SmartOS. They’ll be making their way to an illumos distribution near you some time soon.

Now that we have this, we’re starting to make plans for the future. One idea that Adam had is to make it easy to scope a structure definition to a target process’s data model.

Another big task that we have is to make it easy to get CTF into programs and ideally, make it invisible to anyone using any compiler tool chain on illumos!

With that, happy tracing!

Posted on November 14, 2013 at 11:54 am by rm · Permalink · Comments Closed
In: Miscellaneous

Per-thread caching in libumem

libumem was developed in 2001 by Jeff Bonwick and Jonathan Adams. While the Solaris implementation of malloc(3C) and free(3C) performed adequately for single threaded applications, it did not scale. Drawing on the work that was done to extend the original kernel slab allocator, Jeff and Jonathan brought it to userland in the form of libumem. Since then, libumem has even been brought to other operating systems. libumem offers great performance for multi-threaded applications, though there are a few cases where libumem doesn’t quite perform compared to libc and the allocators found in other operating systems like eglibc. The most common case for this is when you have short-lived small allocations, often less than 200 bytes.

What’s happening?

As part of our work with Voxer, they had uncovered some general performance problems that Brendan and I were digging into. We distilled this down to a small Node.js benchmark that was calculating a lot of md5 sums. As part of narrowing down the problem, I eventually broke out one of Brendan’s flame graphs. Since we had a CPU-bound problem, this allowed us to easily visualize and understand what’s going on. When you throw that under DTrace with jstack(), you get something that looks pretty similar to the following flame graph:

libc flamegraph

There are two main pieces to this flame graph. The first is performing the actual md5 operations. The second is creating and destroying the md5 objects and the initialization associated with that. In drilling down, we found that we were spending comparatively more time trying to handle the allocations. If you look at the flamegraph in detail, you’ll see that when calling malloc and free we’re spending a lot of that time in in the lock portions of libc. libc’s malloc has a global mutex. Using a simple DTrace script like dtrace -n 'pid$target::malloc:entry{ @[tid] = count(); }', we can verify that only one thread is calling into malloc, so we’re grabbing an uncontended lock. One’s next guess might be to try and run this with libumem to see if there is any difference. This gives us the following flame graph:

libumem flamegraph

You can easily spot all of the libumem related allocations because they are a bit more like towers that consist of a series of three functions calls. First to malloc(3C), then umem_alloc(3MALLOC), and finally umem_cache_alloc(3MALLOC). On top of that are additional stacks related to grabbing locks. In umem_cache_alloc there is still a lock that a thread has to grab. Unlike libc’s malloc, this lock is not a global lock. Each cache has a lock per-CPU, which, when combined with magazines allows for highly-parallel allocations. However, we’re only doing mallocs from one thread so this is an uncontested mutex lock. The key takeaway from this is that the uncontested mutex lock can still be problematic. This is also much trickier in user-land where there is a lot more to deal with when grabbing a lock. Compare the kernel implementation with the user-land implementation. One conclusion that you reach from this is that we should do something to get rid of the lock.

When getting rid of mutexes, one might first think of using atomics and trying to rewrite this to be lock-free. But, aside from the additional complexity that rewriting portions of this to be lock-free might induce, that doesn’t solve the fundamental problem: we don’t want to be synchronizing at all. Instead this suggests a different idea that other memory allocators have taken: adding a thread-local cache.

A brief aside: libumem and malloc design

As part of libumem’s design it creates a series of fixed size caches which it uses to handle allocations. These caches are sized from 8 bytes to 128KB, with the difference between caches growing larger with the cache size. If a malloc comes in that is within the range of one of these caches then we use the cache. If the allocation is larger than 128KB then libumem uses a different means to allocate that. For the rest of this entry we’ll only talk about the allocations that are handled by one these caches. For the full details of libumem, I strongly suggest you read the two papers on libumem and the original slab allocator.

Keeping track of your allocations

When you call malloc(3C) you pass in a size, but you don’t need to pass that size back into free(3C). You only need to pass in the buffer. malloc ends up doing work to handle this for you. malloc will actually allocate an extra eight byte tag and prepend that to the buffer. So if you request a 36 bytes, malloc will actually allocate 42 bytes from the system and return you a pointer that starts right after that tag. This tag encodes two pieces of information. The first piece is the size and the second piece is a tag that is encoded with the size. It uses the second field to help detect programs that erroneously write to memory. The structure that it prepends looks like:

typedef struct malloc_data {
	uint32_t malloc_size;
	uint32_t malloc_stat;
} malloc_data_t;

When you call free, libumem grabs that structure, reads the buffer size, and validates the tag. If everything checks out, it releases the entire buffer back to the appropriate cache. If it doesn’t check out, libumem aborts the program.

Per-Thread Caching: High-level

To speed up the allocation and free process, we’re going to change how malloc and free work. When a given thread calls free, instead of releasing the buffer directly back to the cache, we will instead store it with the thread. That way if the thread comes around and requests a buffer that would be satisfied by that cache, it can just take the one that it stored. By creating this per-thread cache, we have a lock-free and contention-free means of servicing allocations. We store these allocations as a linked list and use one list per cache. When we want to add a buffer to the list, we make it the new head. When we remove an entry from the list, we remove the head. If the head is set to NULL then we know that the list is empty. When the list is empty, we simply go ahead and use the normal allocation path. When a thread exits, then all of the buffers in that thread are freed back to the underlying caches.

We don’t want the per-thread cache to end up storing an unbounded amount of memory. That would end up appearing no different from a memory leak. Instead, we have two mechanisms in place to control this.

  1. A cap on the amount of memory each thread may cache.
  2. We only enable this for a subset of umem caches.

By default, each thread is capped at 1 MB of cache. This can be configured on a per process basis using the UMEM_OPTIONS environment variable. Simply set perthread_cache=[size]. The size is in bytes and you can use the common K, M, G, and even T, suffixes. We only support doing this for sixteen caches at this time and we opt to make this the first sixteen caches. If you don’t tune the cache sizes, allocations up to 256 bytes for 32-bit applications and up to 448 bytes for 64-bit applications will be cached.

Finally, when a thread exits, all of the memory in that thread’s cache is released back to the underlying general purpose umem caches.

Another aside: Position Independent Code

Modern libraries are commonly built with Position Independent Code (PIC). The goal of building something PIC is that it can be loaded anywhere in the address space and no additional relocations will need to be performed. This means that all the offsets and addresses within a given library that are for the library itself are relative addresses. The means for doing this for amd64 programs is relatively straightforward. amd64 offers an addressing mode known as RIP-relative. RIP-relative addressing is where you specify an offset relative to the current instruction pointer which is stored in the register %rip. 32-bit i386 programs don’t have RIP-relative addressing, so compilers have to use different tricks to for relative addressing. One of the more common techniques is to use a call +0 instruction to establish a known address. Here is the disassembly of a simple function which happens to call a global function pointer in a library.

amd64-implementation
> testfunc::dis
testfunc:                         movq   +0x1bb39(%rip),%rax      <0x86230>
testfunc+7:                       pushq  %rbp
testfunc+8:                       movq   %rsp,%rbp
testfunc+0xb:                     call   *(%rax)
testfunc+0xd:                     leave
testfunc+0xe:                     ret
i386-implementation
> testfunc::dis
testfunc:                         pushl  %ebp
testfunc+1:                       movl   %esp,%ebp
testfunc+3:                       pushl  %ebx
testfunc+4:                       subl   $0x10,%esp
testfunc+7:                       call   +0x0     <testfunc+0xc>
testfunc+0xc:                     popl   %ebx
testfunc+0xd:                     addl   $0x1a990,%ebx
testfunc+0x13:                    pushl  0x8(%ebp)
testfunc+0x16:                    movl   0x128(%ebx),%eax
testfunc+0x1c:                    call   *(%eax)
testfunc+0x1e:                    addl   $0x10,%esp
testfunc+0x21:                    movl   -0x4(%ebp),%ebx
testfunc+0x24:                    leave
testfunc+0x25:                    ret

Position independent code is still really quite useful, one just has to be aware that they do pay a small price for it. In this case, we’re doing several more loads and stores. When working in intensely performance-sensitive code, those can really add up.

Per-Thread Caching: Implementation

The first problem that we needed to solve for per-thread caching was to figure out how we would store the data for the per-thread caches. While we could have gone with some of the functionality provided by the threading libraries (see threads(5)), that would end up sending us through the Procedure Linkage Table (PLT). Because we are cycle-bumming here our goal is to minimize the number of such calls that we have to make. Instead, we’ve added some storage to the ulwp_t. The ulwp_t is libc’s per-thread data structure. It is the userland equivalent of the kthread_t. We extended the ulwp_t of each thread with the following structure:

typedef struct {
	size_t tm_size;
	void *tm_roots[16];
} tmem_t;

Each entry in the tm_roots array is the head of one of the linked lists that we use to store a set of similarly sized allocations. The tm_size field keeps track of how much data has been stored across all of the linked lists for this thread. Because these structures exist per-thread, there is no need for any synchronization.

Why we need to know about the size of the caches

The set of caches that libumem uses for allocations only exists once umem_init() has finished processing all of the UMEM_OPTIONS environment variables. One of the options can add additional caches and another one can remove caches. It is impossible to know what these caches are at compile time. Given that this is the case, you might ask why do we want to know the size of the caches that libumem creates? Why not just create our own set of sizes that we’re going to use for honoring other allocations?

Failing to use the size of the umem caches would not only cause us to use extra space, but it would also cause us to incur additional fragmentation. Our goal is to be able to reuse these allocations. We can’t have a bucket for every possible allocation size, that would grow quite unwieldy. Let’s say that we used the same bucketing scheme that libumem uses by default. Because we have no way of knowing what cache libumem honored something from, we instead have to assume that the returned allocation is the largest possible size we can use the buffer for. If we make a 65-byte allocation that actually comes from the 80 byte cache, we would instead bucket it in the thread’s 64-byte cache. Effectively, we have to always round an allocation down to the next cache. Items that could satisfy a 64-byte allocation would end up being items that were originally 64-79 bytes large. This is clearly wasteful of our memory.

If you look at the signature of umem_free(3MALLOC) you’ll see that it takes a size argument. This means that it is our responsibility to keep track of what the original size of the allocation was. Normally malloc and free wrap this up in the malloc tags, but since we are reusing buffers, we’ll want to keep track of both the original size and the currently requested size when we reuse it. To do this, we would have to extend the malloc tag structure that we talked about above. While there are some allocations that have extra space for something like this for 64-bit programs, that is not the case for 32-bit programs. To implement it this way, that would require at another eight-byte tag be prepended to every 32-bit malloc and some 64-bit allocations as well.

Obviously this is not an ideal way to go approach the problem. Instead, if we use the cache sizes we don’t have to suffer from either of the above problems. We know that when a umem_alloc comes in, it rounds the allocation size up to the nearest cache to satisfy the request. We leverage that same idea so that when a buffer is freed we put it into the per-thread list that corresponds to the cache that it originally came from. When new requests come in we can use that buffer to satisfy anything that the underlying cache would be able to. Because of this strategy we don’t have to have a second round of fragmentation for our buffers. Similarly, because we know what the cache size is, we don’t have to keep track of the original request size. We know exactly what cache the buffer came from initially.

Following in the footsteps of trapstat and fasttrap

Now that we have opted to know about the cache sizes at run time this means that we have a few approaches we can take to generate the function that handles this per-thread layer. Remember, that we’re here because of performance. We need to cycle-bum and minimize our use of cycles. Loads and stores to and from main memory are much more expensive than simple register arithmetic, comparisons, and small branches. While there is an array of allocations sizes, we don’t want to have to always load each entry from that array. We also have another challenge. We need to avoid doing anything that causes us to use the PLT. We don’t want to end up having a call +0 instruction so we can access 32-bit PIC code. There is one fortunate aspect of the umem caches. Once they are determined at runtime, they never change.

Armed with this information we end up down a different path: we are going to dynamically generate the assembly for our functions. This isn’t the first time this has been done in illumos. For an idea of what this looks like, see trapstat. Our code functions in a very similar way. We have blocks of assembly with holes for addresses and constants. These blocks get copied into executable memory and the appropriate constants get inserted in the correct place. One of the important pieces of this is that we don’t end up calling any other functions. If we detect an error or we can’t satisfy the allocation in the function itself, we end up jumping out to the original malloc and free reusing the same stack frame.

Reaching the assembly generated functions

Once we have generated the machine code, we have the challenge of making it be what applications reach when they call malloc(). This is complicated by the fact that calls to malloc can come in before we create and initialize the new functions as part of the libumem initialization. The standard way you might solve this is with a function pointer that you swap out at some point. However, having that global function pointer causes us to need to address that in a position independent way and adds noticeable overhead. Instead we utilized a small feature of the illumos linker, local auditing, to create a trampoline. Before we get into the details of the auditing library, here’s the data we used to support the decision. We made a small and simple benchmark that just does a fixed number of small mallocs and frees in a loop and compared the differences.

#include <stdlib.h>
#include <stdint.h>
#include <stdio.h>
#include <sys/time.h>

#define MAX     1024 * 1024 * 512

int
main(int argc, char *argv[])
{
        int ii;
        void *f;
        int size = 4;
        hrtime_t start;

        if (argc == 2)
                size = atoi(argv[1]);

        start = gethrtime();
        for (ii = 0; ii < MAX; ii++) {
                f = malloc(4);
                free(f);
        }
        printf("%lld\n", (hrtime_t)(gethrtime() - start));
        return (0);
}

arch libc (ns) libumem (ns) indirect call (ns) audit library (ns)
i386 39833278751 57784034737 14522070966 9215801785
amd64 32470572792 47828105321 9654626131 8980269581

From this data we can see that the audit library technique ended up being a small win on amd64, but for i386, it was a much more substantial win. This all comes down to how the compiler generated PIC code.

Audit libraries are a feature of the illumos linker that allow you to interpose on all the relocations that are being made to and from a given library. For the full details on how audit libraries work consult the Linkers and Libraries guide (one of the best pieces of Sun Documentation) or the linker source itself. We created a local auditing library that allows us to only audit libumem. As part of auditing the relocation for libumem's malloc and free symbols the audit library gives us an opportunity to replace the symbol with one of our own choice. The audit library instead returns the address of a local buffer which contains a jump instruction to the either the actual malloc or free. This installs our simple trampoline.

Later, when umem_init() runs we end up generating the assembly versions of our functions. libumem uses symbols which the auditing library interposes upon to be told where the buffers it should put the generated function are. After both the malloc and free implementations have been successfully generated, it removes the initial jump instruction and atomically replaces it with a five byte nop instruction. We looked at using both the multi-byte nop, five single byte nops, and just zeroing out the jump offset so it would become a jmp +0. Using the same microbenchmark we used earlier, we saw that the multi-byte nop made a noticeable difference.

arch jmp +0 (ns) single-byte nop (ns) multi-byte nop (ns)
i386 9215801785 9434344776 9091563309
amd64 8980269581 8989382774 8562676893

For more details on how this all works and fits together, take a look at the updated libumem big theory statement and pay particular attention to section 8. You may also want to look at the i386 and amd64 specific files which generate the assembly.

Needed linker fixes

There are two changes that are necessary for local auditing to work correctly. Thanks to Bryan who went and made those changes and figured out the way to create this trampoline with the local auditing library.

Understanding per-thread cache utilization

Bryan worked to supply not only the necessary enhancements to the linker, but also supply enhancements to various mdb dcmds to better understand the behavior of the per-thread caching in libumem. ::umastat was enhanced to show both the amount that each thread has used and to show how many allocations each cache has in the per-thread cache (ptc).

> ::umastat
     memory   %   %   %   %   %   %   %   %   %   %   %   %   %   %   %   %   %
tid  cached cap   8  16  32  48  64  80  96 112 128 160 192 224 256 320 384 448
--- ------- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- ---
  1    174K  16   0   6   6   1   4   0   0  18   2  50   0   4   0   1   0   2
  2       0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0
  3    201K  19   0   6   6   2   4   8   1  16   2  43   0   3   0   1   0   1
  4       0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0
  5   62.1K   6   0   8   7   3   9   0   0  13   5  38   0   6   0   2   1   0
  6       0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0 

cache                        buf     buf     buf     buf  memory     alloc alloc
name                        size  in use  in ptc   total  in use   succeed  fail
------------------------- ------ ------- ------- ------- ------- --------- -----
umem_magazine_1               16      33       -     252      4K        35     0
umem_magazine_3               32      36       -     126      4K        39     0
umem_magazine_7               64       0       -       0       0         0     0
umem_magazine_15             128       4       -      31      4K         4     0
umem_magazine_31             256       0       -       0       0         0     0
umem_magazine_47             384       0       -       0       0         0     0
umem_magazine_63             512       0       -       0       0         0     0
umem_magazine_95             768       0       -       0       0         0     0
umem_magazine_143           1152       0       -       0       0         0     0
umem_slab_cache               56      46       -      63      4K        48     0
umem_bufctl_cache             24     153       -     252      8K       155     0
umem_bufctl_audit_cache      192       0       -       0       0         0     0
umem_alloc_8                   8       0       0       0       0         0     0
umem_alloc_16                 16    2192    1827    2268     36K      2192     0
umem_alloc_32                 32    1082     921    1134     36K      1082     0
umem_alloc_48                 48     275     202     336     16K       275     0
umem_alloc_64                 64     487     359     504     32K       487     0
umem_alloc_80                 80     246     234     250     20K       246     0
umem_alloc_96                 96      42      41      42      4K        42     0
umem_alloc_112               112     741     676     756     20K       741     0
umem_alloc_128               128     133     109     155     20K       133     0
umem_alloc_160               160    1425    1274    1425     36K      1425     0
umem_alloc_192               192      11       9      21      4K        11     0
umem_alloc_224               224      83      82      90     20K        83     0
umem_alloc_256               256       8       8      15      4K         8     0
umem_alloc_320               320      24      22      24      8K        24     0
umem_alloc_384               384       7       6      10      4K         7     0
umem_alloc_448               448      20      19      27     12K        20     0
umem_alloc_512               512       1       -      16      8K       138     0
umem_alloc_640               640       0       -      18     12K       130     0
umem_alloc_768               768       0       -      10      8K        87     0
umem_alloc_896               896       0       -      18     16K       114     0
...

In addition, this work inspired Bryan to go and add %H to mdb_printf for human readable sizes. As a part of the support for the enhanced ::umastat, there are also new walkers for the various ptc caches.

Performance of the Per-Thread Caching

The ultimate goal of this was to improve our performance. As part of that we did different testing to make sure that we understood what the impact and effects of this would be. We primarily compared ourselves to the libc malloc implementation and a libumem without our new bits. We also did some comparison to other allocators, notably eglibc. We chose eglibc because that is what the majority of customers coming to us from other systems are using and because it is a good allocator, particularly for small objects.

Tight 4 byte allocation loop

One of the basic things that we wanted to test, inspired in part by some of the behavior we had seen in applications, was to measure what a tight malloc and free loop of a small allocation looked like where we varied the number of threads. Below we included a test where we did this one thread and one where we did it with sixteen threads. The main take away we got from this is that libumem has historically been slow at these compared to a single threaded libc program. The sixteen thread graph shows why we really want to use libumem compared to libc. The graph shows the time per thread. As we can see, libc's single mutex for malloc is rather painful.

4 byte allocations with one thread
4 byte allocations with sixteen threads

Time for all cached allocations

Another thing that we wanted to measure was how our allocation time scaled with the size of the cache. While our assembly is straightforward, it could probably be further optimized. We ran the test with both 32-bit and 64-bit code and the results are below. From the graphs you can see that scale fairly linearly across the caches.

32-bit small allocations
64-bit small allocations

The effects of the per-thread cache on uncacheable allocations

One of the things that we wanted to verify was that the presence of the per-thread caching did not unreasonably degrade the performance of other allocations. To look at this we compared what happened if you used libumem and what happened if you did not. We used pbind to lock the program to a single CPU, measured the time it took to do 1KB sized allocations, and compared the differences. We took that value and divided by the total number of allocations we had performed, 512 M in this case. The end result was that for a given loop of malloc and free, the overhead was 8-10ns. That was within reason for our acceptable overhead.

umem_init time

Another one of the areas we wanted to make sure that we didn't seriously regress was the time it takes umem_init. I've included a coarse graph that was created using DTrace. I simply rigged up something that traced the amount of wall and cpu time umem_init took. We repeated that 100 times and graphed the results. The graph below shows a roughly 50 microsecond increase in the wall and cpu time. In this case, a reasonable increase.

umem_init time

Our Original Flame Graph

The last thing that I want to look at is what the original flame graph now looks like using per-thread-caching. We increased the per-thread cache to 64MB because that allows us to cache the majority of the malloc activity which primarily comes from only one thread. The new flame graph is different from the previous two. The amount of time that we've spent in malloc and free has been minimized and compared to libumem previously, we are no longer three layers deep. In terms of raw improvement, while this normally took around 110 seconds with libc, with per-thread-caching we're down to around 78 seconds. Remember, this is a pure node.js benchmark. To have an improvement to malloc() result in a ~30% win was pretty surprising. Even in dynamic garbage collected languages, the performance of malloc is still important.

ptcumem flamegraph

Wrapping Up

In this entry I've described the high-level design, implementation, and some of the results we've seen from our work on per-thread caching libumem. For some workloads there can be a substantial performance improvement by using per-thread caching. To try it out, grab the upcoming release of SmartOS and either add -lumem to your Makefile or simply try it out by running your application with LD_PRELOAD=libumem.so.

When you link with libumem, per-thread caching is enabled by default with a 1 MB per-thread cache. This value can be tuned via the UMEM_OPTIONS environment variable via UMEM_OPTION=perthread_cache=[size]. For example, to set it to 64 MB, you would do something like: UMEM_OPTIONS=perthread_cache=64M. If you enable any kind of the umem_debug(3MALLOC) facilities then this will be disabled. Similarly if you request nomagazines, this will be disabled.

If you have questions, feel free to ask here or join us on IRC.

Posted on July 16, 2012 at 10:50 am by rm · Permalink · Comments Closed
In: SunOS · Tagged with: , , , ,

mdb tab completion

Last October, the first illumos hack-a-thon took place. Out of that a lot of interesting things were done and have since been integrated into illumos. Two of the more interesting gems were Adam Leventhal and Matt Ahrens adding dtrace -x temporal and Eric Schrock adding the DTrace print() action. Already print() is in the ranks of things where once you have it you really miss it when you don’t. During the hack-a-thon I had the chance to work Matt Amdur. Together we worked on another one of those nice to haves that has finally landed in illumos: tab completion for mdb.

md-what?

For those who have never used it, mdb is the Modular Debugger that comes as a part of illumos and was originally developed for Solaris 8. mdb is primarily used for post-mortem of user and kernel applications and kernel debugger. mdb isn’t a source level debugger, but it works quite well on core dumps from userland, inspects and modifies live kernel state without stopping the system, and provides facilities for traditional debugging where a program is stopped, stepped over, and inspected. mdb replaced adb which came out of AT&T. While mdb isn’t 100% compatible with adb, it does remind you that there’s ‘No algol 68 here’. For the full history, take a look at Mike Shapiro’s talk that he gave at the Brown CS 37th IPP Symposium.

One of the more useful pieces of mdb is its module API which allows other modules to deliver specifically tailored commands and walkers. This is used for everything from the v8 Javascript Engine to understanding cyclics. Between that, pipelines, and other niceties, there isn’t too much else you could want from your debugger.

What’s involved

The work that we’ve done falls into three parts:

Thanks to CTF data in the kernel, we can tab complete everything from walker names, to types and their members. We went and added tab completion to the following dcmds:

Seeing is believing: Tab completion in action

Completing dcmds

> ::pr[tab]
print
printf
project
prov_tab
prtconf
prtusb

Completing walkers

> ::walk ar[tab]
arc_buf_hdr_t
arc_buf_t
> ::walk arc_buf_

Completing types

> ::sizeof struct dt[tab]
struct dtrace_actdesc
struct dtrace_action
struct dtrace_aggbuffer
struct dtrace_aggdesc
struct dtrace_aggkey
struct dtrace_aggregation
struct dtrace_anon
struct dtrace_argdesc
struct dtrace_attribute
struct dtrace_bufdesc
struct dtrace_buffer
struct dtrace_conf
struct dtrace_cred
struct dtrace_difo
struct dtrace_diftype
struct dtrace_difv
struct dtrace_dstate
struct dtrace_dstate_percpu
struct dtrace_dynhash
struct dtrace_dynvar
struct dtrace_ecb
struct dtrace_ecbdesc
struct dtrace_enabling
struct dtrace_eprobedesc
struct dtrace_fmtdesc
struct dtrace_hash
...

Completing members

> ::offsetof zio_t io_v[tab]
io_vd
io_vdev_tree
io_vsd
io_vsd_ops

Walking across types with ::print

> p0::print proc_t p_zone->zone_n[tab]
zone_name
zone_ncpus
zone_ncpus_online
zone_netstack
zone_nlwps
zone_nlwps_ctl
zone_nlwps_lock
zone_nodename
zone_nprocs
zone_nprocs_ctl
zone_nprocs_kstat
zone_ntasks

In addition, just as you can walk across structure (.) and array ([]) dereferences in ::print invocations, you can also do the same with tab completion.

What’s next?

Now that mdb tab completion change is in illumos there’s already some work to add backends to new dcmds including:

What else would you like to see? Let us know in a comment or better yet, go ahead and implement it yourself!

Posted on May 15, 2012 at 11:38 am by rm · Permalink · 2 Comments
In: Miscellaneous

illumos Hardware Compatibility List

One of the challenges when using any Operating System is answering the question ‘Is my hardware supported?’. To track this down, you often have to scour Internet sites, hoping someone else has already asked the question, or do other, more horrible machinations – or ask someone like me. If you’re running on an illumos-based system like SmartOS, OmniOS, or OpenIndiana, this just got a lot easier: I’ve created the list. Better yet, I’ve created a tool to automatically create the list.

The List

illumos now has a hardware compatibility list (HCL) available at http://illumos.org/hcl.

This list contains all the PCI and PCI Express devices that should work. If your PCI device isn’t listed there, don’t fret, it may still work. This list is a first strike at the problem of hardware compatibility, so things like specific motherboards aren’t listed there.

How it’s generated

The great thing about this list is that it’s automatically generated from the source code in illumos itself. Each driver on the system has a manifest that specifies what PCI IDs it supports. We parse each of these manifests and look up the names using the PCI ID Database, using a small library that I wrote. From there, we automatically generate the static web page that can be deployed. Thanks to K. Adam White for his invaluable help to stop me from fumbling around too much with front end web code and the others who have already come in and improved it.

All the code is available on github. The goal for all of this is to eventually be a part of the illumos-gate itself. If you have improvements or want to make the web page more visually appealing, we’d all welcome the contribution.

Posted on May 10, 2012 at 6:00 pm by rm · Permalink · Comments Closed
In: Miscellaneous

Figuring out where you’re longjmp(3c)ing with DTrace

Last Monday was the illumos hack-a-thon. There, I worked with Matt Amdur on adding tab completion support to mdb — the illumos modular debugger. The hack-a-thon was wildly successful and a lot of fun, I hope to put together an entry on the hack-a-thon and give an overview of the projects that were worked on over the course of the next few days. During the hack-a-thon, Matt and I created a working prototype that would complete the types and members using ::print, but there was still some good work for us to do. One of the challenges that we were facing was some unexpected behavior whenever the mdb pager was invoked. We were seeing different actions depending on which actions you took from the pager: continue, quit, or abort.

If you take a look at the source code, you’ll see that sometimes we’ll leave this function by calling longjmp(3c). There’s a lot of different places that we call setjmp(3c) and sigsetjmp(3c) in mdb, so tracking down where we were going just based on looking at the source code would be tricky. So, we want to answer the question, where are we jumping to? There are a few different ways we can do this (and more that aren’t listed):

  1. Inspect the source code
  2. Use mdb to debug mdb
  3. Use the DTrace pid provider and trace a certain number of instructions before we assume we’ve gotten there
  4. Use the DTrace pid provider and look at the jmp_buf to get the address of where we were jumping

Ultimately, I decided to go with option four, knowing that I would have to solve this problem again at some point in the future. The first step is to look at the definition of the jmp_buf definition. For the full definition take a look at setjmp_iso.h. Here’s the snippet that actually defines the type:

     82 #if defined(__i386) || defined(__amd64) || \
     83 	defined(__sparc) || defined(__sparcv9)
     84 #if defined(_LP64) || defined(_I32LPx)
     85 typedef long	jmp_buf[_JBLEN];
     86 #else
     87 typedef int	jmp_buf[_JBLEN];
     88 #endif
     89 #else
     90 #error "ISA not supported"
     91 #endif

Basically, the jmp_buf is just an array where we store some of registers. Unfortunately this isn’t sufficient to figure out where to go. So instead, we need to take a look at the implementation. setjmp is implemented in assembly for the particular architecture. Here it is for x86 and amd64. Now that we have the implementation, let’s figure out what to do. As a heads up, if you’re looking at any of these .s files, the numbers are actually in base 10, which is different from what you get when you look at the mdb output which has them in hex. Let’s take a quick look at the longjmp source for a 32-bit system and dig into what’s going on and how we know what to do:

     73 	ENTRY(longjmp)
     74 	movl	4(%esp),%edx	/ first parameter after return addr
     75 	movl	8(%esp),%eax	/ second parameter
     76 	movl	0(%edx),%ebx	/ restore ebx
     77 	movl	4(%edx),%esi	/ restore esi
     78 	movl	8(%edx),%edi	/ restore edi
     79 	movl	12(%edx),%ebp	/ restore caller's ebp
     80 	movl	16(%edx),%esp	/ restore caller's esp
     81
     82 	movl	24(%edx), %ecx
     83 	test	%ecx, %ecx	/ test flag word
     84 	jz	1f
     85 	xorl	%ecx, %ecx	/ if set, clear ul_siglink
     86 	movl	%ecx, %gs:UL_SIGLINK
     87 1:
     88 	test	%eax,%eax	/ if val != 0
     89 	jnz	1f		/ 	return val
     90 	incl	%eax		/ else return 1
     91 1:
     92 	jmp	*20(%edx)	/ return to caller
     93 	SET_SIZE(longjmp)

The function is pretty well commented, so we can follow along pretty easily. Basically we load the jmp_buf that was passed in into %edx, add 0×14 to that value and then jump to that piece of code. So now we know exactly what the address we’re returning to is. With this in hand, we only have two tasks left: transforming this address into a function and offset, and doing this all with a simple DTrace script. Solving the first problem is actually pretty easy. We can just use the DTrace uaddr function which will translate it into an address and offset for us. The script itself is now an exercise in copyin and arithmetic. Here’s the main part of the script:

/*
 * Given a sigbuf translate that into where the longjmp is taking us.
 *
 * On i386 the address is 0x14 into the jmp_buf.
 * On amd64 the address is 0x38 into the jmp_buf.
 */

pid$1::longjmp:entry
{
        uaddr(curpsinfo->pr_dmodel == PR_MODEL_ILP32 ?
            *(uint32_t *)copyin(arg0 + 0x14, sizeof (uint32_t)) :
            *(uint64_t *)copyin(arg0 + 0x38, sizeof (uint64_t)));
}

Now, if we run this, here’s what we get:

[root@bh1-build2 (bh1) /var/tmp]# dtrace -s longjmp.d $(pgrep -z rm mdb)
dtrace: script 'longjmp.d' matched 1 probe
CPU     ID                    FUNCTION:NAME
  8  69580                    longjmp:entry   mdb`mdb_run+0x38

Now we know exactly where we ended up after the longjmp(), and this will method will work on both 32-bit and 64-bit x86 systems. If you’d like to download the script, you can just download it from here.

Posted on October 29, 2011 at 6:28 pm by rm · Permalink · One Comment
In: DTrace, Miscellaneous · Tagged with: , ,

Visualizing KVM

Last March, Bryan Cantrill and I joined Max Bruning on working towards bringing KVM to illumos. Six months ago we found ourselves looping in x86 real mode and today we’re booting everything from Linux to Plan 9 to Weenix! For a bit more background on how we got there take a gander at Bryan’s entry on KVM on illumos.

For the rest of this entry I’m going to talk about the exciting new analytics we get by integrating DTrace and kstats into KVM. We’ve only scratched the surface of what we can see, but already we’ve integrated several metrics into Cloud Analytics and have gained insight into different areas of guest behavior that the guests themselves haven’t really seen before. While we can never gain the same amount of insight into Virtual Machines (VMs) that we can with a zone, we easily have insight into three main resources of a VM: CPU, disk, and network. Cloud Operators can use these metrics to determine if there is a problem with a VM, determine which VMs are having issues, and what areas of the system are suffering. In addition, we’ve pushed the boundaries of observability by taking advantage of the fact that several components of the hardware stack are virtualized. All in all, we’ve added metrics in the following areas:

NICs and Disks

One of the things that we had to determine early on was how the guests virtual devices would interface with the host. For NICs, this was simple: rather than trying to map a guest’s NIC to a host’s TUN or TAP device; we just used a VNIC, which was introduced into the OS by the Crossbow project. Each guest NIC corresponds directly to a Crossbow VNIC. This allows us to leverage all of the benefits of using a VNIC including anti-spoof and the analytics that already exist. This lets us see the throughput in terms of either bytes or packets that the guest is sending and receiving on a per guest NIC basis.

The story with disks is quite similar. In this case each disk that the guest sees is backed by a zvol from ZFS. This means that guests unknowingly get the benefits of ZFS: data checksums, snapshots and clones, the ease of transfer via zfs send and zfs receive, redundant pooled storage, and a proven reliability. What is more powerful here is the insight that we can provide into the disk operations. We provide two different views of disk activity. The first is based on throughput and the second is based on I/O operations.

The throughput based analytics are a great way to understand the actual disk activity that the VM is doing. Yet the operations view gives us several interesting avenues to drill down into VM activity. The main decompositions are operation type, latency, offset, and size. This gives us insight into how guest filesystems are actually sending activity to disk. As an example, we generated the following screenshot from a guest running Ubuntu on an ext3 filesystem. The guest would loop creating a 3gb file, sleeping for a seconds, reading the file, and deleting the file before beginning again. In the image below we see operations decomposed by operation type and offset. This allows us to see where on disk ext3 is choosing to lay out blocks on the filesystem. The x-axis represents time; each unit is one second. The y-axis shows the virtual disk block number.

ext3 offsets

Hardware Interrupts

Brendan Gregg has been helping us out by diligently measuring our performance, comparing us to both a bare metal Linux system and KVM under Linux. While trying to understand the performance characteristics and ensuring that KVM on illumos didn’t have too many performance pathologies he stumbled across an interesting function in the kvm source code: vmx_inject_irq. This function is called any time a hardware interrupt is injected into the guest. We combined this information with an incredibly valuable idea for heatmaps that Brendan thought up. A heatmap based upon subsecond offset allows us to see when across a given second some action occurred. The x-axis is the same as the previous graph, one second. The y-axis though represents when in the second this item occurred, i.e. where in the 1,000,000 microseconds did this action occur. Take a look at the following image:

subsecond offset by irqs

Here we are visualizing which interrupts occurred in a given second and looking at it based upon when in the second they occur. Each interrupt vector is colored differently. The red represents interrupts caused by the disk controller and yellow by the network controller. The blue is the most interesting: these are timer based interrupts generated by the APIC. Lines that are constant across the horizontal means that these are events that are happening at the same time every second. These represent actions caused by an interval timer, something that
always fires every second. However there are lines that look like a miniature stair; ones that go up at an angle. These represent an application that does work, calls sleep(3C) for an interval, does a little bit of work and sleeps again.

VM Exits

A VM exits when the processor ceases running the guest code and returns to the host kernel to handle something which must be emulated such as memory mapped I/O or accessing one of the emulated devices. One of the ways to increase VM performance is to minimize the number of exits. Early on during our work porting KVM we saw that there were various statistics that were being gathered and exported via debugfs in the Linux KVM code. Here we leveraged kstats and Bryan quickly wrote up kvmstat. kvmstat quickly became an incredibly useful tool for us to easily understand VM behavior. What we’ve done is leverage the kstats which allow us to know which VM, which vCPU, and which of a multitude of reasons the guest exited and add that insight into Cloud Analytics.

vCPU Samples

While working on KVM and reading through the Intel Architecture Manuals I reminded myself of a portion of x86 architecture that is quite important, mainly that the root of the page table is always going to be stored in cr3. Unique values in cr3 represent unique Memory Management Unit (MMU) contexts. Most modern operating systems that run on x86 use a different a different MMU context for each process running on the system and the kernel. Thus if we look at the value of cr3 we get an opaque token that represents something running in the guest.

Brendan had recently been adding metrics to Cloud Analytics based upon using DTrace’s profile provider and combining the gathered data with the subsecond offset heatmaps that we previously discussed. Bryan had added a new variable to D that allowed us to look at the register state of a given running vCPU. To get the value of cr3 that we wanted we could use something along the lines of vmregs[VMX_GUEST_CR3]. When you combine these two we get a heatmap that shows us what is running in the guest. Check out the image below:

vCPU samples by cr3 and subsecond offset

Here, we’ve sampled at a frequency of 99 Hz. We avoid sampling at 100 Hz because that would catch a lot of periodic system activity. We’re looking at a one CPU guest running Ubuntu. The guest is initially idle. Next we start two CPU bound activities, highlighted in blue and yellow. What we can visualize are the scheduling decisions being made by the Linux kernel. To further see what happened, we used renice on one of the processes setting it to a value of 19. You can see the effect in the first image below as the blue process rarely runs in comparison to the yellow. In the second image we experimented with the effects of setting different nice values.

vCPU samples by cr3 and subsecond offset

vCPU samples by cr3 and subsecond offset

These visualizations are quite useful. They let us give someone an idea of what is running in their VM. While it can’t pinpoint it to the exact process, it does let the user understand what the characteristics of their workload are and whether it is a few long lived processes fighting for the CPU, lots of short lived processes coming and going, or something in between. Like the rest of these metrics this lets you understand where in your fleet of VMs the problem may be occurring and help narrow things down to which few VMs should be looked at with native tools.

Conclusion

We’ve only begun to scratch the surface of what we can understand about a virtual machine running under KVM on illumos. Needless to say, this wouldn’t be possible without DTrace and its guarantees of safety for use on production systems and only the overhead of a few NOPs when not in use. As time goes on we’ll be experimenting on what information can help us, operators, and end users better understand their VM’s performance and adding those abilities to Cloud Analytics.

Posted on August 16, 2011 at 11:40 am by rm · Permalink · 2 Comments
In: DTrace · Tagged with: , , ,