diff --git a/docs/EINTR.md b/docs/EINTR.md
new file mode 100644
index 0000000..8d1ab52
--- /dev/null
+++ b/docs/EINTR.md
@@ -0,0 +1,91 @@
+# EINTR
+
+## The problem
+
+If your code is blocked in a system call when a signal needs to be delivered,
+the kernel needs to interrupt that system call. For something like a read(2)
+call where some data has already been read, the call can just return with
+what data it has. (This is one reason why read(2) sometimes returns less data
+than you asked for, even though more data is available. It also explains why
+such behavior is relatively rare, and a cause of bugs.)
+
+But what if read(2) hasn't read any data yet? Or what if you've made some other
+system call, for which there is no equivalent "partial" success, such as
+poll(2)? In poll(2)'s case, there's either something to report (in which
+case the system call would already have returned), or there isn't.
+
+The kernel's solution to this problem is to return failure (-1) and set
+errno to `EINTR`: "interrupted system call".
+
+### Can I just opt out?
+
+Technically, yes. In practice on Android, no. Technically if a signal's
+disposition is set to ignore, the kernel doesn't even have to deliver the
+signal, so your code can just stay blocked in the system call it was already
+making. In practice, though, you can't guarantee that all signals are either
+ignored or will kill your process... Unless you're a small single-threaded
+C program that doesn't use any libraries, you can't realistically make this
+guarantee. If any code has installed a signal handler, you need to cope with
+`EINTR`. And if you're an Android app, the zygote has already installed a whole
+host of signal handlers before your code even starts to run. (And, no, you
+can't ignore them instead, because some of them are critical to how ART works.
+For example: Java `NullPointerException`s are optimized by trapping `SIGSEGV`
+signals so that the code generated by the JIT doesn't have to insert explicit
+null pointer checks.)
+
+### Why don't I see this in Java code?
+
+You won't see this in Java because the decision was taken to hide this issue
+from Java programmers. Basically, all the libraries like `java.io.*` and
+`java.net.*` hide this from you. (The same should be true of `android.*` too,
+so it's worth filing bugs if you find any exceptions that aren't documented!)
+
+### Why doesn't libc do that too?
+
+For most people, things would be easier if libc hid this implementation
+detail. But there are legitimate use cases, and automatically retrying
+would hide those. For example, you might want to use signals and `EINTR`
+to interrupt another thread (in fact, that's how interruption of threads
+doing I/O works in Java behind the scenes!). As usual, C/C++ choose the more
+powerful but more error-prone option.
+
+## The fix
+
+### Easy cases
+
+In most cases, the fix is simple: wrap the system call with the
+`TEMP_FAILURE_RETRY` macro. This is basically a while loop that retries the
+system call as long as the result is -1 and errno is `EINTR`.
+
+So, for example:
+```
+  n = read(fd, buf, buf_size); // BAD!
+  n = TEMP_FAILURE_RETRY(read(fd, buf, buf_size)); // GOOD!
+```
+
+### close(2)
+
+TL;DR: *never* wrap close(2) calls with `TEMP_FAILURE_RETRY`.
+
+The case of close(2) is complicated. POSIX explicitly says that close(2)
+shouldn't close the file descriptor if it returns `EINTR`, but that's *not*
+true on Linux (and thus on Android). See
+[Returning EINTR from close()](https://lwn.net/Articles/576478/)
+for more discussion.
+
+Given that most Android code (and especially "all apps") are multithreaded,
+retrying close(2) is especially dangerous because the file descriptor might
+already have been reused by another thread, so the "retry" succeeds, but
+actually closes a *different* file descriptor belonging to a *different*
+thread.
+
+### Timeouts
+
+System calls with timeouts are the other interesting case where "just wrap
+everything with `TEMP_FAILURE_RETRY()`" doesn't work. Because some amount of
+time will have elapsed, you'll want to recalculate the timeout. Otherwise you
+can end up with your 1 minute timeout being indefinite if you're receiving
+signals at least once per minute, say. In this case you'll want to do
+something like adding an explicit loop around your system call, calculating
+the timeout _inside_ the loop, and using `continue` each time the system call
+fails with `EINTR`.