The method to epoll’s madness

Cindy Sridharan
16 min readOct 30, 2017

My previous post covered the fundamentals of file descriptors as well as some of the most commonly used forms on non-blocking I/O operations on Linux and BSD. I had some people wonder why it didn’t cover epoll at all, but I’d mentioned in the conclusion of that post that epoll is by far the most interesting of all and as such warranted a separate post in its own right.

epoll stands for event poll and is a Linux specific construct. It allows for a process to monitor multiple file descriptors and get notifications when I/O is possible on them. It allows for both edge-triggered as well as level-triggered notifications. Before we look into the bowels of epoll, first let’s explore the syntax.

The syntax of epoll

Unlike poll, epoll itself is not a system call. It’s a kernel data structure that allows a process to multiplex I/O on multiple file descriptors.

This data structure can be created, modified and deleted by three system calls.

1) epoll_create

The epoll instance is created by means of the epoll_create system call, which returns a file descriptor to the epoll instance. The signature of epoll_create is as follows:

#include <sys/epoll.h>
int epoll_create(int size);

The size argument is an indication to the kernel about the number of file descriptors a process wants to monitor, which helps the kernel to decide the size of the epoll instance. Since Linux 2.6.8, this argument is ignored because the epoll data structure dynamically resizes as file descriptors are added or removed from it.

The epoll_create system call returns a file descriptor to the newly created epoll kernel data structure. The calling process can then use this file descriptor to add, remove or modify other file descriptors it wants to monitor for I/O to the epoll instance.

There is another system call epoll_create1 which is defined as follows:

int epoll_create1(int flags);

The flags argument can either be 0 or EPOLL_CLOEXEC.

When set to 0, epoll_create1 behaves the same way as epoll_create.

When the EPOLL_CLOEXEC flag is set, any child process forked by the current process will close the epoll descriptor before it execs, so the child process won’t have access to the epoll instance anymore.

It’s important to note that the file descriptor associated with the epoll instance needs to be released with a close() system call. Multiple processes might hold a descriptor to the same epoll instance, since, for example, a fork without the EPOLL_CLOEXEC flag will duplicate the descriptor to the epoll instance in the child process). When all of these processes have relinquished their descriptor to the epoll instance (by either calling close() or by exiting), the kernel destroys the epoll instance.

2) epoll_ctl

A process can add file descriptors it wants monitored to the epoll instance by calling epoll_ctl. All the file descriptors registered with an epoll instance are collectively called an epoll set or the interest list.

In the above diagram, process 483 has registered file descriptors fd1, fd2, fd3, fd4 and fd5 with the epoll instance. This is the interest list or the epoll set of that particular epoll instance. Subsequently, when any of the file descriptors registered become ready for I/O, then they are considered to be in the ready list.

The ready list is a subset of the interest list.

The signature of the epoll_ctl syscall is as follows:

#include <sys/epoll.h>
int epoll_ctl(int epfd, int op, int fd, struct epoll_event *event);

epfd — is the file descriptor returned by epoll_create which identifies the epoll instance in the kernel.

fd — is the file descriptor we want to add to the epoll list/interest list.

op — refers to the operation to be performed on the file descriptor fd. In general, three operations are supported:

Register fd with the epoll instance (EPOLL_CTL_ADD) and get notified about events that occur on fd
Delete/deregister fd from the epoll instance. This would mean that the process would no longer get any notifications about events on that file descriptor (EPOLL_CTL_DEL). If a file descriptor has been added to multiple epoll instances, then closing it will remove it from all of the epoll interest lists to which it was added.
Modify the events fd is monitoring (EPOLL_CTL_MOD)

event — is a pointer to a structure called epoll_event which stores the event we actually want to monitor fd for.

The first field events of the epoll_event structure is a bitmask that indicates which events fd is being monitored for.

Like so, if fd is a socket, we might want to monitor it for the arrival of new data on the socket buffer (EPOLLIN). We might also want to monitor fd for edge-triggered notifications which is done by OR-ing EPOLLET with EPOLLIN. We might also want to monitor fd for the occurrence of a registered event but only once and stop monitoring fd for subsequent occurrences of that event. This can be accomplished by OR-ing the other flags (EPOLLET, EPOLLIN) we want to set for descriptor fd with the flag for only-once notification delivery EPOLLONESHOT. All possible flags can be found in the man page.

The second field of the epoll_event struct is a union field.

3) epoll_wait

A thread can be notified of events that happened on the epoll set/interest set of an epoll instance by calling the epoll_wait system call, which blocks until any of the descriptors being monitored becomes ready for I/O.

The signature of epoll_wait is as follows:

#include <sys/epoll.h>
int epoll_wait(int epfd, struct epoll_event *evlist, int maxevents, int timeout);

epfd — is the file descriptor returned by epoll_create which identifies the epoll instance in the kernel.

evlist — is an array of epoll_event structures. evlist is allocated by the calling process and when epoll_wait returns, this array is modified to indicate information about the subset of file descriptors in the interest list that are in the ready state (this is called the ready list)

maxevents — is the length of the evlist array

timeout — this argument behaves the same way as it does for poll or select. This value specifies for how long the epoll_wait system call will block:

— when the timeout is set to 0, epoll_wait does not block but returns immediately after checking which file descriptors in the interest list for epfd are ready
— when timeout is set to -1, epoll_wait will block “forever”. When epoll_wait blocks, the kernel can put the process to sleep until epoll_wait returns. epoll_wait will block until 1) one or more descriptors specified in the interest list for epfd become ready or 2) the call is interrupted by a signal handler
— when timeout is set to a non negative and non zero value, then epoll_wait will block until 1) one or more descriptors specified in the interest list for epfd becomes ready or 2) the call is interrupted by a signal handler or 3) the amount of time specified by timeout milliseconds have expired

The return values of epoll_wait are the following:

— if an error (EBADF or EINTR or EFAULT or EINVAL) occurred, then the return code is -1
— if the call timed out before any file descriptor in the interest list became ready, then the return code is 0
— if one or more file descriptors in the interest list became ready, then the return code is a positive integer which indicates the total number of file descriptors in the evlist array. The evlist is then examined to determine which events occurred on which file descriptors.

The gotchas of epoll

To fully understand the nuance behind epoll, it’s important to understand how file descriptors really work. This was explored in my previous post, but it’s worth restating again.

A process references I/O streams with the help of descriptors. Every process maintains a table of file descriptors which it has access to. Every entry in this table has two fields:

— flags controlling the operation of the file descriptor (the only such flag is the close on exec flag)
— a pointer to an underlying kernel data structure we’ll explore in a bit

Descriptors are either created explicitly by system calls like open, pipe, socket and so forth or are inherited from the parent process during a fork. Descriptors are also “duplicated” with a dup/dup2 system call.

Descriptors are released when:

— the process exits
— by calling the close system call
— when a process forks, all the descriptors are “duplicated” in the child process. If any of the descriptors are marked close-on-exec, then after the parent forks but before the child execs, the descriptors in the child marked as close-on-exec are closed and will no longer be available to the child process. The parent can still continue using the descriptor but the child wouldn’t be able to use it once it has exec-ed.

Let us assume in the above example process A has descriptor 3 marked with the close-on-exec flag. If process A forks process B, then immediately after the fork, process A and process B are identical, and as such process B will have “access” to file descriptors 0, 1, 2 and 3.

But since descriptor 3 is marked as close-on-exec, before process B execs, this descriptor will be marked as “inactive”, and process B won’t be able to access it anymore.

To really understand what this means, it becomes important to understand that a descriptor really is just a per process pointer to an underlying kernel data structure called (confusingly) the file description.

The kernel maintains a table of all open file descriptions called the open file table.

Let’s assume fd3 of process A was created as a result of a dup or an fcntl system call on descriptor fd0. Both the original descriptor fd0 and the “duplicated” descriptor fd3 point to the same file description in the kernel.

If process A then forks process B and fd3 is marked with the close-on-exec flag, then the child process B will inherit all of the parent process A’s descriptors but cannot use fd3.

It’s important to note that fd0 in the child process B will also point to the same open file description in the kernel’s open file table.

We have three descriptors — fd0 and fd3 in Process A and fd0 in Process B — that all point to the same underlying kernel open file description. Hold this thought, because this has some important implications for epoll. All other file descriptors in both processes A and B also point to an entry in the open file table, but have been omitted from the diagram.

Note - File descriptions aren’t just shared by two processes when one forks the other. If one process passed a file descriptor to another process over a Unix Domain Socket socket, then the descriptors of both processes again point to the same underlying kernel open file description.

Finally, it becomes important to understand what the inode pointer field of a file description is. But prior to that, it’s important to understand what an inode is.

An inode is file system data structure that contains information about a filesystem object like a file or a directory. This information includes:

— the location of the blocks on disk where the file or directory data is stored
— the attributes of the file or directory
— additional metadata about the file or directory, such as access time, owner, permissions and so forth.

Every file (and directory) in the file system has an inode entry, which is a number that refers to the file. This number is also called the inode number. On many file systems, the maximum number of inodes is capped to a certain value, meaning the total number of files that can be stored on the system is capped too.

There’s an inode table entry on disk that maintains a map of the inode number to the actual inode data structure on disk. Most file systems are accessed via the kernel’s file system driver. This driver uses the inode number to access the information stored in the inode. Thus in order to know the location of a file or any metadata pertaining to the file, the kernel’s file system driver needs to access the inode table.

Let’s assume after process A forks process B, process A has created two more file descriptors fd4 and fd5. These aren’t duplicated in process B.

Let’s assume fd5 is created as a result of process A calling open on file abc.txt for reading. Let us assume process B also calls openon abc.txt but for writing and the file descriptor the open call returns to process B is fd10.

Then process A’s fd5 and process B’s fd10 point to different open file descriptions in the open file table, but they point to the same inode table entry (or in other words, the same file).

This has two very important implications:

— Since fd0 in both process A and process B refer to the same open file description, they share the file offset. This means that if process A advances the file offset (by calling read()or write() or lseek()), then the offset changes for process B as well. This is also applicable to fd3 belonging to process A, since fd3 refers to the same open file description as fd0.
— This is also applicable to modifications made by a file descriptor in one process to an open file status flag ( O_ASYNC, O_NONBLOCK, O_APPEND). So if process B sets fd0 to the non blocking mode by setting the O_NONBLOCKflag via the fcntlsystem call, then descriptors fd0 and fd3 belonging to process A will also start observing non-blocking behavior.

The bowels of epoll

Let us assume a process A has two open file descriptors fd0 and fd1, that have two open file descriptions in the open file table. Let is assume both these file descriptions point to different inodes.

epoll_create creates a new inode entry (the epoll instance) as well as an open file description for it in the kernel, and returns to the calling process a file descriptor (fd9) to this open file description.

When we use epoll_ctl to add a file descriptor (say fd0) to the epoll instance’s interest list, we’re actually fd0’s underlying file description to the epoll instance’s interest list.

Thus the epoll instance actually monitors the underlying file description, and not the per process file descriptor. This has some interesting implications.

— If process A forks a child process B, then B inherits all of A’s descriptors, including fd9, the epoll descriptor. However, process B’s descriptors fd0, fd1 and fd9 still refer to the same underlying kernel data structures. Process B’s epoll descriptor (fd9) shares the same interest list with process A.

If after the fork, if process A creates creates a new descriptor fd8 (non-duplicated in process B) to its epoll interest list via epoll_ctl, then it’s not just process A that gets notifications about events on fd8 when calling epoll_wait()

If process B calls epoll_wait(), then process B gets the notification about fd8 (which belongs to process A and wasn’t duplicated during the fork) as well. This is also applicable when the epoll file descriptor is duplicated by means if a call to dup/dup2 or if the epoll file descriptor is passed to another process over a Unix Domain Socket.

Let’s assume process B opens the file pointed to by fd8 with a new open call, and gets a new file descriptor (fd15) as a result. Let’s now assume process A closes fd8. One would assume that since process A has closed fd8, it will no longer get notifications about events on fd8 when calling epoll_wait. This, however, isn’t the case, since the interest list monitors the open file description. Since fd15 points to the same description as fd8 (since they are both the same underlying file), process A gets notifications about events on fd15. It’s safe to say that once a file descriptor has been registered by a process with the epoll instance, then the process will continue getting notifications about events on the descriptor even if it closes the descriptor, so long as the underlying open file description is still referenced by at least one other descriptor (belonging to the same or a different process).

Why epoll is more performant that select and poll

As stated in the previous post, the cost of select/poll is O(N), which means when N is very large (think of a web server handling tens of thousands of mostly sleepy clients), every time select/poll is called, even if there might only be a small number of events that actually occurred, the kernel still needs to scan every descriptor in the list.

Since epoll monitors the underlying file description, every time the open file description becomes ready for I/O, the kernel adds it to the ready list without waiting for a process to call epoll_wait to do this. When a process does call epoll_wait, then at that time the kernel doesn’t have to do any additional work to respond to the call, but instead returns all the information about the ready list it’s been maintaining all along.

Furthermore, with every call to select/poll requires passing the kernel the information about the descriptors we want to monitored. This is obvious from the signature to both calls. The kernel returns the information about all the file descriptors passed in which the process again needs to examine (by scanning all the descriptors) to find out which ones are ready for I/O.

int poll(struct pollfd *fds, nfds_t nfds, int timeout);int select(int nfds, fd_set *readfds, fd_set *writefds, fd_set *exceptfds, struct timeval *timeout);

With epoll, once we add the file descriptors to the epoll instance’s interest list using the epoll_ctl call, then when we call epoll_waitin the future, we don’t need to subsequently pass the file descriptors whose readiness information we wish to find out. The kernel again only returns back information about those descriptors which are ready for I/O, as opposed to the select/poll model where the kernel returns information about every descriptor passed in.

As a result, the cost of epoll is O(number of events that have occurred) and not O(number of descriptors being monitored) as was the case with select/poll.

Edge triggered epoll

By default, epoll provides level-triggered notifications. Every call to epoll_wait only returns the subset of file descriptors belonging to the interest list that are ready.

So if we have four file descriptors (fd1, fd2, fd3 and fd4) registered, and only two (fd2 and fd3) are ready at the time of calling epoll_wait, then only information about these two descriptors are returned.

It’s also interesting to note that in the default level-triggered case, the nature of the descriptors (blocking versus non-blocking) in epoll’s interest won’t really affect the result of an epoll_waitcall, since epoll only ever updates its ready list when the underlying open file description becomes ready.

Sometimes we might just want to find the status of any descriptor (say fd1, for example) in the interest list, irrespective of whether it’s ready or not. epoll allows us to find out whether I/O is possible on any particular file descriptor (even if it’s not ready at the time of calling epoll_wait) by means of supporting edge-triggered notifications. If we want information about whether there has been any I/O activity on a file descriptor since the previous call to epoll_wait (or since the descriptor was opened, if there was no previous epoll_wait call made by the process), we can get edge-triggered notifications by ORing the EPOLLET flag while calling epoll_ctl while registering a file descriptor with the epoll instance.

Perhaps it becomes more helpful to see this in action in code from a real project where a file descriptor is being registered with an epoll instance with epoll_ctl where the EPOLLET flag is ORed along with some other flags.

function Poller:register(fd, r, w)
local ev = self.ev[0]
ev.events = bit.bor(C.EPOLLET, C.EPOLLERR, C.EPOLLHUP)
if r then
ev.events = bit.bor(ev.events, C.EPOLLIN)
end
if w then
ev.events = bit.bor(ev.events, C.EPOLLOUT)
end
ev.data.u64 = fd
local rc = C.epoll_ctl(self.fd, C.EPOLL_CTL_ADD, fd, ev)
if rc < 0 then errors.get(rc):abort() end
end

Perhaps an illustrative example can better help understand how edge-triggered notifications work with epoll. Let’s use the example used previously, where a process has registered four descriptors with the epoll instance. Let’s assume that fd3 is a socket.

Let’s assume that at time t1, an input byte stream arrives on the socket referenced by fd3.

At t0, input arrives on the socket.

Let’s now assume that at time t4, the process calls epoll_wait().

If at time t4, file descriptors fd2 and fd3 are ready, then the epoll_wait call reports fd2 and fd3 as ready.

At time t4, the process calls epoll_wait

Let’s assume that the process calls epoll_wait again at time t6. Let’s assume fd1 is ready. Let’s also assume that no input arrived on the socket referenced by fd3 between times t4 and t6.

In the level-triggered case, a call to epoll_wait will return fd1 to the process, since fd1 is the only descriptor that is ready. However in the edge-triggered case, this call will block, since no new data has arrived on the socket referenced by fd3 between times t4 and t6.

At time t6, the process calls epoll_wait again

Conclusion

This post aimed to capture the “method” part. In order to understand the “madness” wrecked by these semantics of epoll, a good reference would be the following two blog posts:

epoll is fundamentally broken parts 1 and 2.

--

--

Cindy Sridharan

@copyconstruct on Twitter. views expressed on this blog are solely mine, not those of present or past employers.