Как изменить конфиг nginx

NGINX Basics

Go to the Table of Contents or What’s next? section.

• ≡ NGINX Basics
  • Directories and files
  • Commands
  • Processes
    • CPU pinning
    • Shutdown of worker processes
  • Configuration syntax
    • Comments
    • End of lines
    • Variables, Strings, and Quotes
    • Directives, Blocks, and Contexts
    • External files
    • Measurement units
    • Regular expressions with PCRE
    • Enable syntax highlighting
  • Connection processing
    • Event-Driven architecture
    • Multiple processes
    • Simultaneous connections
    • HTTP Keep-Alive connections
    • sendfile, tcp_nodelay, and tcp_nopush
  • Request processing stages
  • Server blocks logic
    • Handle incoming connections
    • Matching location
    • rewrite vs return
    • URL redirections
    • try_files directive
    • if, break, and set
    • root vs alias
    • internal directive
    • External and internal redirects
    • allow and deny
    • uri vs request_uri
  • Compression and decompression
    • What is the best NGINX compression gzip level?
  • Hash tables
    • Server names hash table
  • Log files
    • Conditional logging
    • Manually log rotation
    • Error log severity levels
    • How to log the start time of a request?
    • How to log the HTTP request body?
    • NGINX upstream variables returns 2 values
  • Reverse proxy
    • Passing requests
    • Trailing slashes
    • Passing headers to the backend
      • Importance of the Host header
      • Redirects and X-Forwarded-Proto
      • A warning about the X-Forwarded-For
      • Improve extensibility with Forwarded
    • Response headers
  • Load balancing algorithms
    • Backend parameters
    • Upstream servers with SSL
    • Round Robin
    • Weighted Round Robin
    • Least Connections
    • Weighted Least Connections
    • IP Hash
    • Generic Hash
    • Other methods
  • Rate limiting
    • Variables
    • Directives, keys, and zones
    • Burst and nodelay parameters
  • NAXSI Web Application Firewall
  • OWASP ModSecurity Core Rule Set (CRS)
  • Core modules
    • ngx_http_geo_module
  • 3rd party modules
    • ngx_set_misc
    • ngx_http_geoip_module

Directories and files

If you compile NGINX with default parameters all files and directories are available from /usr/local/nginx location.

For upstream NGINX packaging paths can be as follows (it depends on the type of system/distribution):

• /etc/nginx — is the default configuration root for the NGINX service

  • other locations: /usr/local/etc/nginx, /usr/local/nginx/conf

• /etc/nginx/nginx.conf — is the default configuration entry point used by the NGINX services, includes the top-level http block and all other configuration contexts and files

  • other locations: /usr/local/etc/nginx/nginx.conf, /usr/local/nginx/conf/nginx.conf

• /usr/share/nginx — is the default root directory for requests, contains html directory and basic static files

  • other locations: html/ in root directory

• /var/log/nginx — is the default log (access and error log) location for NGINX

  • other locations: logs/ in root directory

• /var/cache/nginx — is the default temporary files location for NGINX

  • other locations: /var/lib/nginx

• /etc/nginx/conf — contains custom/vhosts configuration files

  • other locations: /etc/nginx/conf.d, /etc/nginx/sites-enabled (I can’t stand this debian/apache-like convention)

• /var/run/nginx — contains information about NGINX process(es)

  • other locations: /usr/local/nginx/logs, logs/ in root directory

See also Installation and Compile-Time Options — Files and Permissions.

Commands

🔖 Use reload option to change configurations on the fly — Base Rules — P2

• nginx -h — shows the help
• nginx -v — shows the NGINX version
• nginx -V — shows the extended information about NGINX: version, build parameters, and configuration arguments
• nginx -t — tests the NGINX configuration
• nginx -c <filename> — sets configuration file (default: /etc/nginx/nginx.conf)
• nginx -p <directory> — sets prefix path (default: /etc/nginx/)
• nginx -T — tests the NGINX configuration and prints the validated configuration on the screen
• nginx -s <signal> — sends a signal to the NGINX master process:
  • stop — discontinues the NGINX process immediately
  • quit — stops the NGINX process after it finishes processing
    inflight requests
  • reload — reloads the configuration without stopping processes
  • reopen — instructs NGINX to reopen log files
• nginx -g <directive> — sets global directives out of configuration file

Some useful snippets for management of the NGINX daemon:

• testing configuration:

  /usr/sbin/nginx -t -c /etc/nginx/nginx.conf
  /usr/sbin/nginx -t -q -g 'daemon on; master_process on;' # ; echo $?
  
  /usr/local/etc/rc.d/nginx status

• starting daemon:

  /usr/sbin/nginx -g 'daemon on; master_process on;'
  
  service nginx start
  systemctl start nginx
  
  /usr/local/etc/rc.d/nginx start
  
  # You can also start NGINX from start-stop-daemon script:
  /sbin/start-stop-daemon --quiet --start --exec /usr/sbin/nginx --background --retry QUIT/5 --pidfile /run/nginx.pid

• stopping daemon:

  # graceful shutdown (waiting for the worker processes to finish serving current requests)
  /usr/sbin/nginx -s quit
  # fast shutdown (kill connections immediately)
  /usr/sbin/nginx -s stop
  
  service nginx stop
  systemctl stop nginx
  
  /usr/local/etc/rc.d/nginx stop
  
  # You can also stop NGINX from start-stop-daemon script:
  /sbin/start-stop-daemon --quiet --stop --retry QUIT/5 --pidfile /run/nginx.pid

• reloading daemon:

  /usr/sbin/nginx -g 'daemon on; master_process on;' -s reload
  
  service nginx reload
  systemctl reload nginx
  
  /usr/local/etc/rc.d/nginx reload
  
  kill -HUP $(cat /var/run/nginx.pid)
  kill -HUP $(pgrep -f "nginx: master")

• restarting daemon:

  service nginx restart
  systemctl restart nginx
  
  /usr/local/etc/rc.d/nginx restart

Something about testing configuration:

You cannot test half-baked configurations. For example, you defined a server section for your domain in a separate file. Any attempt to test such a file will throw errors. The file has to be complete in all respects.

Configuration syntax

🔖 Organising Nginx configuration — Base Rules — P2
🔖 Format, prettify and indent your Nginx code — Base Rules — P2

NGINX uses a micro programming language in the configuration files. This language’s design is heavily influenced by Perl and Bourne Shell. Configuration syntax, formatting and definitions follow a so-called C-style convention. For me, NGINX configuration has a simple and very transparent structure.

Comments

NGINX configuration files don’t support comment blocks, they only accept # at the beginning of a line for a comment.

End of lines

Lines containing directives must end with a semicolon (;), otherwise NGINX will fail to load the configuration and report an error.

Variables, Strings, and Quotes

Variables start with $ and that get set automaticaly for each request. The ability to set variables at runtime and control logic flow based on them is part of the rewrite module and not a general feature of NGINX. By default, we cannot modify built-in variables like $host or $request_uri.

There are some directives that do not support variables, e.g. access_log (is really the exception because can contain variables with restrictions) or error_log. Variables probably can’t be (and shouldn’t be because they are evaluated in the run-time during the processing of each request and rather costly compared to plain static configuration) declared anywhere, with very few exceptions: root directive can contains variables, server_name directive only allows strict $hostname built-in value as a variable-like notation (but it’s more like a magic constant). If you use variables in if context, you can only set them in if conditions (and maybe rewrite directives). Don’t try to use them elsewhere.

To assign value to the variable you should use a set directive:

See if, break, and set section to learn more about variables.

Some interesting things about variables:

Make sure to read the agentzh’s Nginx Tutorials — it’s about NGINX tips & tricks. This guy is a NGINX Guru and creator of the OpenResty. In these tutorials he describes, amongst other things, variables in great detail. I also recommend nginx built-in variables post.

• the most variables in NGINX only exist at runtime, not during configuration time
• the scope of variables spreads out all over configuration
• variable assignment occurs when requests are actually being served
• variable have exactly the same lifetime as the corresponding request
• each request does have its own version of all those variables’ containers (different containers values)
• requests do not interfere with each other even if they are referencing a variable with the same name
• the assignment operation is only performed in requests that access location

Strings may be inputted without quotes unless they include blank spaces, semicolons or curly braces, then they need to be escaped with backslashes or enclosed in single/double quotes.

Quotes are required for values which are containing space(s) and/or some other special characters, otherwise NGINX will not recognize them. You can either quote or -escape some special characters like " " or ";" in strings (characters that would make the meaning of a statement ambiguous). So the following instructions are the same:

# 1)
add_header My-Header "nginx web server;";

# 2)
add_header My-Header nginx web server;;

Variables in quoted strings are expanded normally unless the $ is escaped.

Directives, Blocks, and Contexts

Read this great article about the NGINX configuration inheritance model by Martin Fjordvald.

Configuration options are called directives. We have four types of directives:

• standard directive — one value per context:

• array directive — multiple values per context:

  error_log /var/log/nginx/localhost/localhost-error.log warn;

• action directive — something which does not just configure:

  rewrite ^(.*)$ /msie/$1 break;

• try_files directive:

  try_files $uri $uri/ /test/index.html;

Valid directives begin with a variable name and then state an argument or series of arguments separated by spaces.

Directives are organised into groups known as blocks or contexts. Generally, context is a block directive that can have other directives inside braces. It appears to be organised in a tree-like structure, defined by sets of brackets — { and }.

The curly braces actually denote a new configuration context.

As a general rule, if a directive is valid in multiple nested scopes, a declaration in a broader context will be passed on to any child contexts as default values. The children contexts can override these values at will.

Directives placed in the configuration file outside of any contexts are considered to be in the global/main context.

Special attention should be paid to some strange behavior associated with some directives. For more information please see Set the HTTP headers with add_header and proxy_*_header directives properly rule.

Directives can only be used in the contexts that they were designed for. NGINX will error out on reading a configuration file with directives that are declared in the wrong context.

If you want to review all directives see alphabetical index of directives.

Contexts can be layered within one another (a level of inheritance). Their structure looks like this:

Global/Main Context
        |
        |
        +-----» Events Context
        |
        |
        +-----» HTTP Context
        |          |
        |          |
        |          +-----» Server Context
        |          |          |
        |          |          |
        |          |          +-----» Location Context
        |          |
        |          |
        |          +-----» Upstream Context
        |
        |
        +-----» Mail Context

The most important contexts are shown in the following description. These will be the ones that you will be dealing with for the most part:

• global — contains global configuration directives; is used to set the settings for NGINX globally and is the only context that is not surrounded by curly braces

• events — configuration for the events module; is used to set global options for connection processing; contains directives that affect connection processing are specified

• http — controls all the aspects of working with the HTTP module and holds directives for handling HTTP and HTTPS traffic; directives in this context can be grouped into:

  • HTTP client directives
  • HTTP file I/O directives
  • HTTP hash directives
  • HTTP socket directives

• server — defines virtual host settings and describes a logical separation of a set of resources associated with a particular domain or IP address

• location — define directives to handle client request and indicates a URI that comes either from the client or from an internal redirect

• upstream — define a pool of back-end servers that NGINX can proxy the request; commonly used for defining either a web server cluster for load balancing

NGINX also provides other contexts (e.g. used for mapping) such as:

• map — is used to set the value of a variable depending on the value of another variable. It provides a mapping of one variable’s values to determine what the second variable should be set to

• geo — is used to specify a mapping. However, this mapping is specifically used to categorize client IP addresses. It sets the value of a variable depending on the connecting IP address

• types — is used to map MIME types to the file extensions that should be associated with them

• if — provide conditional processing of directives defined within, execute the instructions contained if a given test returns true

• limit_except — is used to restrict the use of certain HTTP methods within a location context

Look also at the graphic below. It presents the most important contexts with reference to the configuration:

For HTTP, NGINX lookup starts from the http block, then through one or more server blocks, followed by the location block(s).

External files

include directive may appear inside any contexts to perform conditional inclusion. It attaching another file, or files matching the specified mask:

include /etc/nginx/proxy.conf;

# or:
include /etc/nginx/conf/*.conf;

You cannot use variables in NGINX config file includes. This is because includes are processed before any variables are evaluated.

See also this:

Variables should not be used as template macros. Variables are evaluated in the run-time during the processing of each request, so they are rather costly compared to plain static configuration. Using variables to store static strings is also a bad idea. Instead, a macro expansion and «include» directives should be used to generate configs more easily and it can be done with the external tools, e.g. sed + make or any other common template mechanism.

Measurement units

It is recommended to always specify a suffix for the sake of clarity and consistency.

Sizes can be specified in:

• without a suffix: Bytes
• k or K: Kilobytes
• m or M: Megabytes
• g or G: Gigabytes

Time intervals can be specified in:

• without a suffix: Seconds
• ms: Milliseconds
• s: Seconds
• m: Minutes
• h: Hours
• d: Days
• w: Weeks
• M: Months (30 days)
• y: Years (365 days)

proxy_read_timeout 20; # =20s, default

Some of the time intervals can be specified only with a seconds resolution. You should also remember about this:

Multiple units can be combined in a single value by specifying them in the order from the most to the least significant, and optionally separated by whitespace. For example, 1h 30m specifies the same time as 90m or 5400s.

Regular expressions with PCRE

🔖 Enable PCRE JIT to speed up processing of regular expressions — Performance — P2

Before start reading next chapters you should know what regular expressions are and how they works (they are not a black magic really). I recommend two great and short write-ups about regular expressions created by Jonny Fox:

• Regex tutorial — A quick cheatsheet by examples
• Regex cookbook — Top 10 Most wanted regex

Why? Regular expressions can be used in both the server_name and location (also in other) directives, and sometimes you must have a great skills of reading them. I think you should create the most readable regular expressions that do not become spaghetti code — impossible to debug and maintain.

NGINX uses the PCRE library to perform complex manipulations with your location blocks and use the powerful rewrite directive. To use a regular expression for string matching, it first needs to be compiled, which is usually done at the configuration phase.

You can also enable pcre_jit to dynamic translation during execution (at run time) rather than prior to execution. This option can improve performance, however, in some cases pcre_jit may have a negative effect. So, before enabling it, I recommend you to read this great document: PCRE Performance Project.

Below is also something interesting about regular expressions and PCRE:

• Learn PCRE in Y minutes
• PCRE Regex Cheatsheet
• Regular Expression Cheat Sheet — PCRE
• Regex cheatsheet
• Regular expressions in Perl
• Regexp Security Cheatsheet
• A regex cheatsheet for all those regex haters (and lovers)

You can also use external tools for testing regular expressions. For more please see online tools chapter.

If you’re good at it, check these very nice and brainstorming regex challenges:

• RegexGolf
• Regex Crossword

Enable syntax highlighting

vi/vim

# 1) Download vim plugin for NGINX:

# Official NGINX vim plugin:
mkdir -p ~/.vim/syntax/

wget "http://www.vim.org/scripts/download_script.php?src_id=19394" -O ~/.vim/syntax/nginx.vim

# Improved NGINX vim plugin (incl. syntax highlighting) with Pathogen:
mkdir -p ~/.vim/{autoload,bundle}/

curl -LSso ~/.vim/autoload/pathogen.vim https://tpo.pe/pathogen.vim
echo -en "nexecute pathogen#infect()n" >> ~/.vimrc

git clone https://github.com/chr4/nginx.vim ~/.vim/bundle/nginx.vim

# 2) Set location of NGINX config files:
cat > ~/.vim/filetype.vim << __EOF__
au BufRead,BufNewFile /etc/nginx/*,/etc/nginx/conf.d/*,/usr/local/nginx/conf/*,*/conf/nginx.conf if &ft == '' | setfiletype nginx | endif
__EOF__

It may be interesting for you: Highlight insecure SSL configuration in Vim.

Sublime Text

Install cabal — system for building and packaging Haskell libraries and programs (on Ubuntu):

add-apt-repository -y ppa:hvr/ghc
apt-get update

apt-get install -y cabal-install-1.22 ghc-7.10.2

# Add this to the main configuration file of your shell:
export PATH=$HOME/.cabal/bin:/opt/cabal/1.22/bin:/opt/ghc/7.10.2/bin:$PATH
source $HOME/.<shellrc>

cabal update

• nginx-lint:

  git clone https://github.com/temoto/nginx-lint
  
  cd nginx-lint && cabal install --global

• sublime-nginx + SublimeLinter-contrib-nginx-lint:

  Bring up the Command Palette and type install. Among the commands you should see Package Control: Install Package. Type nginx to install sublime-nginx and after that do the above again for install SublimeLinter-contrib-nginx-lint: type SublimeLinter-contrib-nginx-lint.

Processes

🔖 Adjust worker processes — Performance — P3
🔖 Improve debugging by disable daemon, master process, and all workers except one — Debugging — P4

NGINX has one master process and one or more worker processes. It has also cache loader and cache manager processes but only if you enable caching.

The main purposes of the master process is to read and evaluate configuration files, as well as maintain the worker processes (respawn when a worker dies), handle signals, notify workers, opens log files, and, of course binding to ports.

Master process should be started as root user, because this will allow NGINX to open sockets below 1024 (it needs to be able to listen on port 80 for HTTP and 443 for HTTPS).

To defines the number of worker processes you should set worker_processes directive.

The worker processes do the actual processing of requests and get commands from master process. They runs in an event loop (registering events and responding when one occurs), handle network connections, read and write content to disk, and communicate with upstream servers. These are spawned by the master process, and the user and group will as specified (unprivileged).

The worker processes spend most of the time just sleeping and waiting for new events (they are in S state in top).

The following signals can be sent to the NGINX master process:

There’s no need to control the worker processes yourself. However, they support some signals too:

CPU pinning

Moreover, it is important to mention about worker_cpu_affinity directive (it’s only supported on GNU/Linux). CPU affinity is used to control which CPUs NGINX utilizes for individual worker processes. By default, worker processes are not bound to any specific CPUs. What’s more, system might schedule all worker processes to run on the same CPU which may not be efficient enough.

CPU affinity is represented as a bitmask (given in hexadecimal), with the lowest order bit corresponding to the first logical CPU and the highest order bit corresponding to the last logical CPU.

Here you will find an amazing explanation of this. There is a worker_cpu_affinity configuration generator for NGINX. After all, I would recommend to let the OS scheduler to do the work because there is no reason to ever set it up during normal operation.

Shutdown of worker processes

This should come in useful if you want to tweak NGINX’s shutdown process, particularly if other servers or load balancers are relying upon predictable restart times or if it takes a long time to close worker processes.

The worker_shutdown_timeout directive configures a timeout to be used when gracefully shutting down worker processes. When the timer expires, NGINX will try to close all the connections currently open to facilitate shutdown.

NGINX’s Maxim Dounin explains:

The worker_shutdown_timeout directive is not expected to delay shutdown if there are no active connections. It was introduced to limit possible time spent in shutdown, that is, to ensure fast enough shutdown even if there are active connections.

When a worker process enters the «exiting» state, it does a few things:

• mark itself as an exiting process
• set a shutdown timer, if worker_shutdown_timeout is defined
• close listening sockets
• close idle connections

Then, if the shutdown timer was set, after the worker_shutdown_timeout interval, all connections are closed.

By default, NGINX to wait for and process additional data from a client before fully closing a connection, but only if heuristics suggests that a client may be sending more data.

Sometimes, you can see nginx: worker process is shutting down in your log file. The problem occurs when reloading the configuration — where NGINX usually exits the existing worker processes gracefully, but at times, it takes hours to close these processes. Every config reload may dropping a zombie workers, permanently eating up all of your system’s memory. In this case, fast shutdown of worker processes might be a solution.

In addition, setting worker_shutdown_timeout also solve the issue:

worker_shutdown_timeout 60s;

Test connection timeouts and how long your request is processed by a server, next adjust the worker_shutdown_timeout value to these values. 60 seconds is a value with a solid supply and nothing valid should last longer than that.

In my experience, if you have multiple workers in a shutting down state, maybe you should first look at the loaded modules that may cause problems with hanging worker processes.

Connection processing

NGINX supports a variety of connection processing methods which depends on the platform used.

In general there are four types of event multiplexing:

• select — is anachronism and not recommended but installed on all platforms as a fallback
• poll — is anachronism and not recommended

And the most efficient implementations of non-blocking I/O:

• epoll — recommend if you’re using GNU/Linux
• kqueue — recommend if you’re using BSD (it is technically superior to epoll)

The select method can be enabled or disabled using the --with-select_module or --without-select_module configuration parameter. Similarly, the poll method can be enabled or disabled using the --with-poll_module or --without-poll_module configuration parameter.

epoll is an efficient method of processing connections available on Linux 2.6+. kqueue is an efficient method of processing connections available on FreeBSD 4.1+, OpenBSD 2.9+, and NetBSD 2.0+.

There is normally no need to specify it explicitly, because NGINX will by default use the most efficient method. But if you want to set this:

There are also great resources (also makes comparisons) about them:

• Kqueue: A generic and scalable event notification facility
• poll vs select vs event-based
• select/poll/epoll: practical difference for system architects
• Scalable Event Multiplexing: epoll vs. kqueue
• Async IO on Linux: select, poll, and epoll
• A brief history of select(2)
• Select is fundamentally broken
• Epoll is fundamentally broken
• I/O Multiplexing using epoll and kqueue System Calls
• Benchmarking BSD and Linux
• The C10K problem

Look also at libevent benchmark (read about libevent – an event notification library):

^{This infographic comes from daemonforums — An interesting benchmark (kqueue vs. epoll).}

You may also view why big players uses NGINX on FreeBSD instead of on GNU/Linux:

• FreeBSD NGINX Performance
• Why did Netflix use NGINX and FreeBSD to build their own CDN?

NGINX means connections as follows (the following status information is provided by ngx_http_stub_status_module):

• Active connections — the current number of active (open) client connections including waiting connections and connections to backends

  • accepts — the total number of accepted client connections
  • handled — the total number of handled connections. Generally, the parameter value is the same as accepts unless some resource limits have been reached (for example, the worker_connections limit)
  • requests — the total number of client requests

• Reading — the current number of connections where NGINX is reading the request header

• Writing — the current number of connections where NGINX is writing the response back to the client (reads request body, processes request, or writes response to a client)

• Waiting — the current number of idle client connections waiting for a request, i.e. connection still opened waiting for either a new request, or the keepalive expiration (actually it is Active — (Reading + Writing))

  Waiting connections those are keepalive connections. They are usually not a problem but if you want to reduce them set the lower value of the keepalive_timeout directive.

Be sure to recommend to read this:

Writing connections counter increasing might indicate one of the following:

• crashed or killed worker processes. This is unlikely in your case though, as this would also result in other values growing as well, notably Waiting
• a real socket leak somewhere. These usually results in sockets in CLOSE_WAIT state (in a waiting state for the FIN packet terminating the connection), try looking at netstat output without grep -v CLOSE_WAIT filter. Leaked sockets are reported by NGINX during graceful shutdown of a worker process (for example, after a configuration reload) — if there are any leaked sockets, NGINX will write open socket ... left in connection ... alerts to the error log

To further investigate things, please do the following:

• upgrade to the latest mainline versions, without any 3rd party modules, and check if you are able to reproduce the issue
• try disabling HTTP/2 to see if it fixes the issue
• check if you are seeing open socket ... left in connection ... (socket leaks) alerts on configuration reload

See also Debugging socket leaks (from this handbook).

Event-Driven architecture

Thread Pools in NGINX Boost Performance 9x! — this official article is an amazing explanation about thread pools and generally about handling connections. I also recommend Inside NGINX: How We Designed for Performance & Scale. Both are really great.

NGINX uses Event-Driven architecture which heavily relies on Non-Blocking I/O. One advantage of non-blocking/asynchronous operations is that you can maximize the usage of a single CPU as well as memory because is that your thread can continue it’s work in parallel. The end result is that even as load increases, memory and CPU usage remain manageable.

There is a perfectly good and brief summary about non-blocking I/O and multi-threaded blocking I/O by Werner Henze. I also recommend asynchronous vs non-blocking by Daniel Earwicker.

Take a look at this simple drawing:

^{This infographic comes from Kansas State Polytechnic website.}

Blocking I/O system calls (a) do not return until the I/O is complete. Nonblocking I/O system calls return immediately. The process is later notified when the I/O is complete.

There are forms of I/O and examples of POSIX functions:

Look also what the official documentation says about it:

It’s well known that NGINX uses an asynchronous, event‑driven approach to handling connections. This means that instead of creating another dedicated process or thread for each request (like servers with a traditional architecture), it handles multiple connections and requests in one worker process. To achieve this, NGINX works with sockets in a non‑blocking mode and uses efficient methods such as epoll and kqueue.

Because the number of full‑weight processes is small (usually only one per CPU core) and constant, much less memory is consumed and CPU cycles aren’t wasted on task switching. The advantages of such an approach are well‑known through the example of NGINX itself. It successfully handles millions of simultaneous requests and scales very well.

I must not forget to mention here about Non-Blocking and 3rd party modules (also from official documentation):

Unfortunately, many third‑party modules use blocking calls, and users (and sometimes even the developers of the modules) aren’t aware of the drawbacks. Blocking operations can ruin NGINX performance and must be avoided at all costs.

To handle concurrent requests with a single worker process NGINX uses the reactor design pattern. Basically, it’s a single-threaded but it can fork several processes to utilize multiple cores.

However, NGINX is not a single threaded application. Each of worker processes is single-threaded and can handle thousands of concurrent connections. Workers are used to get request parallelism across multiple cores. When a request blocks, that worker will work on another request.

NGINX does not create a new process/thread for each connection/requests but it starts several worker threads during start. It does this asynchronously with one thread, rather than using multi-threaded programming (it uses an event loop with asynchronous I/O).

That way, the I/O and network operations are not a very big bottleneck (remember that your CPU would spend a lot of time waiting for your network interfaces, for example). This results from the fact that NGINX only use one thread to service all requests. When requests arrive at the server, they are serviced one at a time. However, when the code serviced needs other thing to do it sends the callback to the other queue and the main thread will continue running (it doesn’t wait).

Now you see why NGINX can handle a large amount of requests perfectly well (and without any problems).

For more information take a look at following resources:

• Asynchronous, Non-Blocking I/O
• Asynchronous programming. Blocking I/O and non-blocking I/O
• Blocking I/O and non-blocking I/O
• Non-blocking I/O
• About High Concurrency, NGINX architecture and internals
• A little holiday present: 10,000 reqs/sec with Nginx!
• Nginx vs Apache: Is it fast, if yes, why?
• How is Nginx handling its requests in terms of tasks or threading?
• Why nginx is faster than Apache, and why you needn’t necessarily care
• How we scaled nginx and saved the world 54 years every day

Finally, look at these great preview:

^{Both infographic comes from Inside NGINX: How We Designed for Performance & Scale.}

Multiple processes

NGINX uses only asynchronous I/O, which makes blocking a non-issue. The only reason NGINX uses multiple processes is to make full use of multi-core, multi-CPU, and hyper-threading systems. NGINX requires only enough worker processes to get the full benefit of symmetric multiprocessing (SMP).

From official documentation:

The NGINX configuration recommended in most cases — running one worker process per CPU core — makes the most efficient use of hardware resources.

NGINX uses a custom event loop which was designed specifically for NGINX — all connections are processed in a highly efficient run-loop in a limited number of single-threaded processes called workers. Worker processes accept new requests from a shared listen socket and execute a loop. There’s no specialized distribution of connections to the workers in NGINX; this work is done by the OS kernel mechanisms which notifies a workers.

Upon startup, an initial set of listening sockets is created. workers then continuously accept, read from and write to the sockets while processing HTTP requests and responses. — from The Architecture of Open Source Applications — NGINX.

Multiplexing works by using a loop to increment through a program chunk by chunk operating on one piece of data/new connection/whatever per connection/object per loop iteration. It is all based on events multiplexing like epoll() or kqueue(). Within each worker NGINX can handle many thousands of concurrent connections and requests per second.

See Nginx Internals presentation as a lot of great stuff about the internals of the NGINX.

NGINX does not fork a process or thread per connection (like Apache) so memory usage is very conservative and extremely efficient in the vast majority of cases. NGINX is a faster and consumes less memory than Apache and performs very well under load. It is also very friendly for CPU because there’s no ongoing create-destroy pattern for processes or threads.

Finally and in summary:

• uses Non-Blocking «Event-Driven» architecture
• uses the single-threaded reactor pattern to handle concurrent requests
• uses highly efficient loop for connection processing
• is not a single threaded application because it starts multiple worker processes (to handle multiple connections and requests) during start

Simultaneous connections

Okay, so how many simultaneous connections can be processed by NGINX?

worker_processes * worker_connections = max connections

According to this: if you are running 4 worker processes with 4,096 worker connections per worker, you will be able to serve 16,384 connections. Of course, these are the NGINX settings limited by the kernel (number of connections, number of open files, or number of processes).

At this point, I would like to mention about Understanding socket and port in TCP. It is a great and short explanation. I also recommend to read Theoretical maximum number of open TCP connections that a modern Linux box can have.

I’ve seen some admins does directly translate the sum of worker_processes and worker_connections into the number of clients that can be served simultaneously. In my opinion, it is a mistake because certain of clients (e.g. browsers which have different values for this) opens a number of parallel connections (see this to confirm my words). Clients typically establish 4-8 TCP connections so that they can download resources in parallel (to download various components that compose a web page, for example, images, scripts, and so on). This increases the effective bandwidth and reduces latency.

That is a HTTP/1.1 limit (6-8) of concurrent HTTP calls. The best solution to improve performance (without upgrade the hardware and use cache at the middle (e.g. CDN, Varnish)) is using HTTP/2 (RFC 7540 ^[IETF]) instead of HTTP/1.1.

HTTP/2 multiplex many HTTP requests on a single connection. When HTTP/1.1 has a limit of 6-8 roughly, HTTP/2 does not have a standard limit but say: «It is recommended that this value (SETTINGS_MAX_CONCURRENT_STREAMS) be no smaller than 100» (RFC 7540). That number is better than 6-8.

Additionally, you must know that the worker_connections directive includes all connections per worker (e.g. connection structures are used for listen sockets, internal control sockets between NGINX processes, connections with proxied servers, and for upstream connections), not only incoming connections from clients.

Be aware that every worker connection (in the sleeping state) needs 256 bytes of memory, so you can increase it easily.

The number of connections is especially limited by the maximum number of open files (RLIMIT_NOFILE) on your system (you can read about file descriptors and file handlers on this great explanation). The reason is that the operating system needs memory to manage each open file, and memory is a limited resource. This limitation only affects the limits for the current process. The limits of the current process are bequeathed to children processes too, but each process has a separate count.

To change the limit of the maximum file descriptors (that can be opened by a single worker process) you can also edit the worker_rlimit_nofile directive. With this, NGINX provides very powerful dynamic configuration capabilities with no service restarts.

The number of file descriptors is not the only one limitation of the number of connections — remember also about the kernel network (TCP/IP stack) parameters and the maximum number of processes.

I don’t like this piece of the NGINX documentation. Maybe I’m missing something but it says the worker_rlimit_nofile is a limit on the maximum number of open files for worker processes. I believe it is associated to a single worker process.

If you set RLIMIT_NOFILE to 25,000 and worker_rlimit_nofile to 12,000, NGINX sets (only for workers) the maximum open files limit as a worker_rlimit_nofile. But the master process will have a set value of RLIMIT_NOFILE. Default value of worker_rlimit_nofile directive is none so by default NGINX sets the initial value of maximum open files from the system limits.

# On GNU/Linux (or /usr/lib/systemd/system/nginx.service):
grep "LimitNOFILE" /lib/systemd/system/nginx.service
LimitNOFILE=5000

grep "worker_rlimit_nofile" /etc/nginx/nginx.conf
worker_rlimit_nofile 256;

   PID       SOFT HARD
 24430       5000 5000
 24431        256 256
 24432        256 256
 24433        256 256
 24434        256 256

# To check fds on FreeBSD:
sysctl kern.maxfiles kern.maxfilesperproc kern.openfiles
kern.maxfiles: 64305
kern.maxfilesperproc: 57870
kern.openfiles: 143

This is also controlled by the OS because the worker is not the only process running on the server. It would be very bad if your workers used up all of the file descriptors available to all processes, don’t set your limits so that is possible.

In my opinion, relying on the RLIMIT_NOFILE (and alternatives on other systems) than worker_rlimit_nofile value is more understandable and predictable. To be honest, it doesn’t really matter which method is used to set, but you should keep a constant eye on the priority of the limits.

If you don’t set the worker_rlimit_nofile directive manually, then the OS settings will determine how many file descriptors can be used by NGINX.

I think that the chance of running out of file descriptors is minimal, but it might be a big problem on a high traffic websites.

Ok, so how many fds are opens by NGINX?

• one file handler for the client’s active connection
• one file handler for the proxied connection (that will open a socket handling these requests to remote or local host/process)
• one file handler for opening file (e.g. static file)
• other file handlers for internal connections, shared libraries, log files, and sockets

Also important is:

NGINX can use up to two file descriptors per full-fledged connection.

Look also at these diagrams:

• 1 file handler for connection with client and 1 file handler for static file being served by NGINX:

  # 1 connection, 2 file handlers
  
                       +-----------------+
  +----------+         |                 |
  |          |    1    |                 |
  |  CLIENT <---------------> NGINX      |
  |          |         |        ^        |
  +----------+         |        |        |
                       |      2 |        |
                       |        |        |
                       |        |        |
                       | +------v------+ |
                       | | STATIC FILE | |
                       | +-------------+ |
                       +-----------------+
  

• 1 file handler for connection with client and 1 file handler for a open socket to the remote or local host/process:

  # 2 connections, 2 file handlers
  
                       +-----------------+
  +----------+         |                 |         +-----------+
  |          |    1    |                 |    2    |           |
  |  CLIENT <---------------> NGINX <---------------> BACKEND  |
  |          |         |                 |         |           |
  +----------+         |                 |         +-----------+
                       +-----------------+
  

• 2 file handlers for two simultaneous connections from the same client (1, 4), 1 file handler for connection with other client (3), 2 file handlers for static files (2, 5), and 1 file handler for a open socket to the remote or local host/process (6), so in total it is 6 file descriptors:

  # 4 connections, 6 file handlers
  
                    4
        +-----------------------+
        |              +--------|--------+
  +-----v----+         |        |        |
  |          |    1    |        v        |  6
  |  CLIENT <-----+---------> NGINX <---------------+
  |          |    |    |        ^        |    +-----v-----+
  +----------+    |    |        |        |    |           |
                3 |    |      2 | 5      |    |  BACKEND  |
  +----------+    |    |        |        |    |           |
  |          |    |    |        |        |    +-----------+
  |  CLIENT  <----+    | +------v------+ |
  |          |         | | STATIC FILE | |
  +----------+         | +-------------+ |
                       +-----------------+
  

In the first two examples: we can take that NGINX needs 2 file handlers for full-fledged connection (but still uses 2 worker connections). In the third example NGINX can take still 2 file handlers for every full-fledged connection (also if client uses parallel connections).

So, to conclude, I think that the correct value of worker_rlimit_nofile per all connections of worker should be greater than worker_connections.

In my opinion, the safe value of worker_rlimit_nofile (and system limits) is:

# 1 file handler for 1 connection:
worker_connections + (shared libs, log files, event pool, etc.) = worker_rlimit_nofile

# 2 file handlers for 1 connection:
(worker_connections * 2) + (shared libs, log files, event pool, etc.) = worker_rlimit_nofile

That is probably how many files can be opened by each worker and should have a value greater than to the number of connections per worker (according to the above formula).

In the most articles and tutorials we can see that this parameter has a value similar to the maximum number (or even more) of all open files by the NGINX. If we assume that this parameter applies to each worker separately these values are altogether excessive.

However, after a deeper reflection they are rational because they allow one worker to use all the file descriptors so that they are not confined to other workers if something happens to them. Remember though that we are still limited by the connections per worker. May I remind you that any connection opens at least one file.

So, moving on, the maximum number of open files by the NGINX should be:

(worker_processes * worker_connections * 2) + (shared libs, log files, event pool, etc.) = max open files

To serve 16,384 connections by all workers (4,096 connections for each worker), and bearing in mind about the other handlers used by NGINX, a reasonably value of max files handlers in this case may be 35,000. I think it’s more than enough.

Given the above to change/improve the limitations you should:

1. Edit the maximum, total, global number of file descriptors the kernel will allocate before choking (this step is optional, I think you should change this only for a very very high traffic):

   # Find out the system-wide maximum number of file handles:
   sysctl fs.file-max
   
   # Shows the current number of all file descriptors in kernel memory:
   #   first value:  <allocated file handles>
   #  second value:  <unused-but-allocated file handles>
   #   third value:  <the system-wide maximum number of file handles> # fs.file-max
   sysctl fs.file-nr
   
   # Set it manually and temporarily:
   sysctl -w fs.file-max=150000
   
   # Set it permanently:
   echo "fs.file-max = 150000" > /etc/sysctl.d/99-fs.conf
   
   # And load new values of kernel parameters:
   sysctl -p       # for /etc/sysctl.conf
   sysctl --system # for /etc/sysctl.conf and all of the system configuration files

2. Edit the system-wide value of the maximum file descriptor number that can be opened by a single process:

   • for non-systemd systems:

     # Set the maximum number of file descriptors for the users logged in via PAM:
     #   /etc/security/limits.conf
     nginx       soft    nofile    35000
     nginx       hard    nofile    35000

   • for systemd systems:

     # Set the maximum number (hard limit) of file descriptors for the services started via systemd:
     #   /etc/systemd/system.conf          - global config (default values for all units)
     #   /etc/systemd/user.conf            - this specifies further per-user restrictions
     #   /lib/systemd/system/nginx.service - default unit for the NGINX service
     #   /etc/systemd/system/nginx.service - for your own instance of the NGINX service
     [Service]
     # ...
     LimitNOFILE=35000
     
     # Reload a unit file and restart the NGINX service:
     systemctl daemon-reload && systemct restart nginx

3. Adjusts the system limit on number of open files for the NGINX worker. The maximum value can not be greater than LimitNOFILE (in this example: 35,000). You can change it at any time:

   # Set the limit for file descriptors for a single worker process (change it as needed):
   #   nginx.conf within the main context
   worker_rlimit_nofile 10000;
   
   # You need to reload the NGINX service:
   nginx -s reload

To show the current hard and soft limits applying to the NGINX processes (with nofile, LimitNOFILE, or worker_rlimit_nofile):

for _pid in $(pgrep -f "nginx: [master,worker]") ; do

  echo -en "$_pid "
  grep "Max open files" /proc/${_pid}/limits | awk '{print $4" "$5}'

done | xargs printf '%6s %10st%sn%6s %10st%sn' "PID" "SOFT" "HARD"

or use the following:

# To determine the OS limits imposed on a process, read the file /proc/$pid/limits.
# $pid corresponds to the PID of the process:
for _pid in $(pgrep -f "nginx: [master,worker]") ; do

  echo -en ">>> $_pid\n"
  cat /proc/$_pid/limits

done

To list the current open file descriptors for each NGINX process:

for _pid in $(pgrep -f "nginx: [master,worker]") ; do

  _fds=$(find /proc/${_pid}/fd/*)
  _fds_num=$(echo "$_fds" | wc -l)

  echo -en "nn##### PID: $_pid ($_fds_num fds) #####nn"

  # List all files from the proc/{pid}/fd directory:
  echo -en "$_fdsnn"

  # List all open files (log files, memory mapped files, libs):
  lsof -as -p $_pid | awk '{if(NR>1)print}'

done

You should also remember about the following rules:

• worker_rlimit_nofile serves to dynamically change the maximum file descriptors the NGINX worker processes can handle, which is typically defined with the system’s soft limit (ulimit -Sn)

• worker_rlimit_nofile works only at the process level, it’s limited to the system’s hard limit (ulimit -Hn)

• if you have SELinux enabled, you will need to run setsebool -P httpd_setrlimit 1 so that NGINX has permissions to set its rlimit. To diagnose SELinux denials and attempts you can use sealert -a /var/log/audit/audit.log, or audit2why and audit2allow tools

To sum up this example:

• each of the NGINX processes (master + workers) have the ability to create up to 35,000 files
• for all workers, the maximum number of file descriptors is 140,000 (LimitNOFILE per worker)
• for each worker, the initial/current number of file descriptors is 10,000 (worker_rlimit_nofile)

nginx: master process         = LimitNOFILE (35,000)
  _ nginx: worker process    = LimitNOFILE (35,000), worker_rlimit_nofile (10,000)
  _ nginx: worker process    = LimitNOFILE (35,000), worker_rlimit_nofile (10,000)
  _ nginx: worker process    = LimitNOFILE (35,000), worker_rlimit_nofile (10,000)
  _ nginx: worker process    = LimitNOFILE (35,000), worker_rlimit_nofile (10,000)

                              = master (35,000), all workers:
                                                 - 140,000 by LimitNOFILE
                                                 - 40,000 by worker_rlimit_nofile

Look also at this great article about Optimizing Nginx for High Traffic Loads.

HTTP Keep-Alive connections

🔖 Activate the cache for connections to upstream servers — Performance — P2

Before starting this section I recommend to read the following articles:

• HTTP Keepalive Connections and Web Performance
• Optimizing HTTP: Keep-alive and Pipelining
• Evolution of HTTP — HTTP/0.9, HTTP/1.0, HTTP/1.1, Keep-Alive, Upgrade, and HTTPS

The original model of HTTP, and the default one in HTTP/1.0, is short-lived connections. Each HTTP request is completed on its own connection; this means a TCP handshake happens before each HTTP request, and these are serialized. The client creates a new TCP connection for each transaction (and the connection is torn down after the transaction completes).

HTTP Keep-Alive connection or persistent connection is the idea of using a single TCP connection to send and receive multiple HTTP requests/responses (Keep Alive’s work between requests), as opposed to opening a new connection for every single request/response pair.

When using keep alive the browser does not have to make multiple connections (keep in mind that establishing connections is expensive) but uses the already established connection and controls how long that stays active/open. So, the keep alive is a way to reduce the overhead of creating the connection, as, most of the time, a user will navigate through the site etc. (plus the multiple requests from a single page, to download css, javascript, images etc.).

It takes a 3-way handshake to establish a TCP connection, so, when there is a perceivable latency between the client and the server, keepalive would greatly speed things up by reusing existing connections.

This mechanism hold open the TCP connection between the client and the server after an HTTP transaction has completed. It’s important because NGINX needs to close connections from time to time, even if you configure NGINX to allow infinite keep alive timeouts and a huge amount of acceptable requests per connection, to return results and as well errors and success messages.

Persistent connection model keeps connections opened between successive requests, reducing the time needed to open new connections. The HTTP pipelining model goes one step further, by sending several successive requests without even waiting for an answer, reducing much of the latency in the network.

^{This infographic comes from Mozilla MDN — Connection management in HTTP/1.x.}

However, at present, browsers are not using pipelined HTTP requests. For more information please see Why is pipelining disabled in modern browsers?.

Look also at this example that shows how a Keep-Alive header could be used:

 Client                        Proxy                         Server
   |                             |                              |
   +- Keep-Alive: timeout=600 -->|                              |
   |  Connection: Keep-Alive     |                              |
   |                             +- Keep-Alive: timeout=1200 -->|
   |                             |  Connection: Keep-Alive      |
   |                             |                              |
   |                             |<-- Keep-Alive: timeout=300 --+
   |                             |    Connection: Keep-Alive    |
   |<- Keep-Alive: timeout=5000 -+                              |
   |    Connection: Keep-Alive   |                              |
   |                             |                              |

NGINX official documentation says:

All connections are independently negotiated. The client indicates a timeout of 600 seconds (10 minutes), but the proxy is only prepared to retain the connection for at least 120 seconds (2 minutes). On the link between proxy and server, the proxy requests a timeout of 1200 seconds and the server reduces this to 300 seconds. As this example shows, the timeout policies maintained by the proxy are different for each connection. Each connection hop is independent.

Keepalive connections reduce overhead, especially when SSL/TLS is in use but they also have drawbacks; even when idling they consume server resources, and under heavy load, DoS attacks can be conducted. In such cases, using non-persistent connections, which are closed as soon as they are idle, can provide better performance. So, Keep-Alives will improve SSL/TLS performance by quite a big deal if clients are doing multiple requests but if you don’t have the resources to handle them then they kill your servers.

NGINX closes keepalive connections when the worker_connections limit is reached (connections are kept in the cache till the origin server closes them).

To better understand how Keep-Alive works, please see amazing explanation by Barry Pollard.

NGINX provides the two layers to enable Keep-Alive:

Client layer

• the maximum number of keepalive requests a client can make over a given connection, which means a client can make e.g. 256 successfull requests inside one keepalive connection:

  # Default: 100
  keepalive_requests 256;

• server will close connection after this time. A higher number may be required when there is a larger amount of traffic to ensure there is no frequent TCP connection re-initiated. If you set it lower, you are not utilizing keep-alives on most of your requests slowing down client:

  # Default: 75s
  keepalive_timeout 10s;
  
  # Or tell the browser when it should close the connection by adding an optional second timeout
  # in the header sent to the browser (some browsers do not care about the header):
  keepalive_timeout 10s 25s;

  Increase this to allow the keepalive connection to stay open longer, resulting in faster subsequent requests. However, setting this too high will result in the waste of resources (mainly memory) as the connection will remain open even if there is no traffic, potentially: significantly affecting performance. I think this should be as close to your average response time as possible. You could also decrease little by little the timeout (75s -> 50s, then later 25s…) and see how the server behaves.

Upstream layer

• the number of idle keepalive connections that remain open for each worker process. The connections parameter sets the maximum number of idle keepalive connections to upstream servers that are preserved in the cache of each worker process (when this number is exceeded, the least recently used connections are closed):

  # Default: disable
  keepalive 32;

NGINX, by default, only talks on HTTP/1.0 to the upstream servers. To keep TCP connection alive both upstream section and origin server should be configured to not finalise the connection.

Please keep in mind that keepalive is a feature of HTTP 1.1, NGINX uses HTTP 1.0 per default for upstreams.

Connection won’t be reused by default because keepalive in the upstream section means no keepalive (each time you can see TCP stream number increases per every request to origin server).

HTTP keepalive enabled in NGINX upstream servers reduces latency thus improves performance and it reduces the possibility that the NGINX runs out of ephemeral ports.

The connections parameter should be set to a number small enough to let upstream servers process new incoming connections as well.

Update your upstream configuration to use keepalive:

upstream bk_x8080 {

  ...

  # Sets the maximum number of idle keepalive connections to upstream servers
  # that are preserved in the cache of each worker process.
  keepalive 16;

}

And enable the HTTP/1.1 protocol in all upstream requests:

server {

  ...

  location / {

    # Default is HTTP/1, keepalive is only enabled in HTTP/1.1:
    proxy_http_version 1.1;
    # Remove the Connection header if the client sends it,
    # it could be "close" to close a keepalive connection:
    proxy_set_header Connection "";

    proxy_pass http://bk_x8080;

  }

}

...

}

There are two basic cases when keeping connections alive is really beneficial:

• fast backends, which produce responses is a very short time, comparable to a TCP handshake
• distant backends, when a TCP handshake takes a long time, comparable to a backend response time

Look at the test:

• without keepalive for upstream:

wrk -c 500 -t 6 -d 60s -R 15000 -H "Host: example.com" https://example.com/
Running 1m test @ https://example.com/
  6 threads and 500 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    24.13s    10.68s   49.55s    59.06%
    Req/Sec   679.21     42.44   786.00     78.95%
  228421 requests in 1.00m, 77.98MB read
  Socket errors: connect 0, read 0, write 0, timeout 1152
  Non-2xx or 3xx responses: 4
Requests/sec:   3806.96
Transfer/sec:      1.30MB

• with keepalive for upstream:

wrk -c 500 -t 6 -d 60s -R 15000 -H "Host: example.com" https://example.com/
Running 1m test @ https://example.com/
  6 threads and 500 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    23.40s     9.53s   47.25s    60.67%
    Req/Sec     0.86k    50.19     0.94k    60.00%
  294148 requests in 1.00m, 100.41MB read
  Socket errors: connect 0, read 0, write 0, timeout 380
Requests/sec:   4902.24
Transfer/sec:      1.67MB

sendfile, tcp_nodelay, and tcp_nopush

Before you start reading please review:

• Nginx optimization, understanding SENDFILE, TCP_NODELAY and TCP_NOPUSH
• Nginx Tutorial #2: Performance

As you’re making these changes, keep careful watch on your network traffic and see how each tweak impacts congestion.

sendfile

By default, NGINX handles file transmission itself and copies the file into the buffer before sending it. Enabling the sendfile directive eliminates the step of copying the data into the buffer and enables direct copying data from one file descriptor to another.

Normally, when a file needs to be sent, the following steps are required:

• malloc — allocate a local buffer for storing object data
• read — retrieve and copy the object into the local buffer
• write — copy the object from the local buffer into the socket buffer

Look at this great explanation (from Nginx Tutorial #2: Performance):

This involves two context switches (read, write) which make a second copy of the same object unnecessary. As you may see, it is not the optimal way. Thankfully, there is another system call that improves sending files, and it’s called (surprise, surprise!): sendfile(2). This call retrieves an object to the file cache, and passes the pointers (without copying the whole object) straight to the socket descriptor. Netflix states that using sendfile(2) increased the network throughput from 6Gbps to 30Gbps.

When a file is transferred by a process, the kernel first buffers the data and then sends the data to the process buffers. The process, in turn, sends the data to the destination.

NGINX employs a solution that uses the sendfile system call to perform a zero-copy data flow from disk to socket and saves context switching from userspace on read/write. sendfile tell how NGINX buffers or reads the file (trying to stuff the contents directly into the network slot, or buffer its contents first).

This method is an improved method of data transfer, in which data is copied between file descriptors within the OS kernel space, that is, without transferring data to the application buffers. No additional buffers or data copies are required, and the data never leaves the kernel memory address space.

In my opinion enabling this really won’t make any difference unless NGINX is reading from something which can be mapped into the virtual memory space like a file (i.e. the data is in the cache). But please… do not let me influence you — you should in the first place be keeping an eye on this document: Optimizing TLS for High–Bandwidth Applications in FreeBSD ^[pdf].

By default NGINX disable the use of sendfile:

# http, server, location, if in location contexts
# To turn on sendfile (my recommendation):
sendfile on;

# To turn off sendfile:
sendfile off;     # default

Look also at sendfile_max_chunk directive. NGINX documentation say:

When set to a non-zero value, limits the amount of data that can be transferred in a single sendfile() call. Without the limit, one fast connection may seize the worker process entirely.

On fast local connection sendfile() in Linux may send tens of megabytes per one syscall blocking other connections. sendfile_max_chunk allows to limit the maximum size per one sendfile() operation. So, with this NGINX can reduce the maximum time spent in blocking sendfile() calls, since NGINX won’t try to send the whole file at once, but will do it in chunks. For example:

sendfile on;
sendfile_max_chunk 512k;

tcp_nodelay

I recommend to read The Caveats of TCP_NODELAY and Rethinking the TCP Nagle Algorithm ^[pdf]. These great papers describes very interesting topics about TCP_NODELAY and TCP_NOPUSH.

tcp_nodelay is used to manage Nagle’s algorithm which is one mechanism for improving TCP efficiency by reducing the number of small packets sent over the network. If you set tcp_nodelay on;, NGINX adds the TCP_NODELAY options when opening a new socket.

The option only affects keep-alive connections. Otherwise there is 100ms delay when NGINX sends response tail in the last incomplete TCP packet. Additionally, it is enabled on SSL connections, for unbuffered proxying, and for WebSocket proxying.

Maybe you should think about enabling Nagle’s algorithm (tcp_nodelay off;) but it really depends on what is your specific workload and dominant traffic patterns on a service. tcp_nodelay on; is more reasonable for the modern web, the whole delay business of TCP was reasonable for terminals. Typically LANs have less issues with traffic congestion as compared to the WANs. The Nagle algorithm is most effective if TCP/IP traffic is generated sporadically by user input, not by applications using stream oriented protocols like a HTTP traffic.

So, for me, the recipe is simple:

• bulk sends or HTTP traffic
• applications that require lower latency
• non-interactive type of traffic

There is no need for using Nagle’s algorithm.

You should also know the Nagle’s algorithm author’s interesting comment:

If you’re doing bulk file transfers, you never hit that problem. If you’re sending enough data to fill up outgoing buffers, there’s no delay. If you send all the data and close the TCP connection, there’s no delay after the last packet. If you do send, reply, send, reply, there’s no delay. If you do bulk sends, there’s no delay. If you do send, send, reply, there’s a delay.

The real problem is ACK delays. The 200ms «ACK delay» timer is a bad idea that someone at Berkeley stuck into BSD around 1985 because they didn’t really understand the problem. A delayed ACK is a bet that there will be a reply from the application level within 200ms. TCP continues to use delayed ACKs even if it’s losing that bet every time.

I think if you are dealing with non-interactive type of traffic or bulk transfers such as HTTP/web traffic then enabling TCP_NODELAY to disable Nagle’s algorithm may be useful (is the default behavior of the NGINX). This is especially relevant if you’re running applications or environments that only sometimes have highly interactive traffic and chatty protocols.

By default NGINX enable the use of TCP_NODELAY option:

# http, server, location contexts
# To turn on tcp_nodelay and at the same time to disable Nagle’s algorithm
# (my recommendation, unless you turn tcp_nopush on):
tcp_nodelay on;   # default

# To turn off tcp_nodelay and at the same time to enable Nagle’s algorithm:
tcp_nodelay off;

tcp_nopush

This option is only available if you are using sendfile (NGINX uses tcp_nopush for requests served with sendfile). It causes NGINX to attempt to send its HTTP response head in one packet, instead of using partial frames. This is useful for prepending headers before calling sendfile, or for throughput optimization.

Normally, using tcp_nopush along with sendfile is very good. However, there are some cases where it can slow down things (specially from cache systems), so, run your own tests and find if it’s useful in that way.

tcp_nopush enables TCP_CORK (more specifically, the TCP_NOPUSH socket option on FreeBSD or the TCP_CORK socket option on Linux) which aggressively accumulates data and which tells TCP to wait for the application to remove the cork before sending any packets.

If TCP_NOPUSH/TCP_CORK (are not the same!) is enabled in a socket, it will not send data until the buffer fills to a fixed limit (allows application to control building of packet, e.g pack a packet with full HTTP response). To read more about it and get into the details of this option please read TCP_CORK: More than you ever wanted to know.

Once, I read that tcp_nopush is opposite to tcp_nodelay. I don’t agree with that because, as I understand it, the first one aggregates data based on buffer pressure instead whereas Nagle’s algorithm aggregates data while waiting for a return ACK, which the latter option disables.

It may appear that tcp_nopush and tcp_nodelay are mutually exclusive but if all directives are turned on, NGINX manages them very wisely:

• ensure packages are full before sending them to the client
• for the last packet, tcp_nopush will be removed, allowing TCP to send it immediately, without the 200ms delay

And let’s also remember (take a look at Tony Finch notes — this guy developed a kernel patch for FreeBSD which makes TCP_NOPUSH work like TCP_CORK):

• on Linux, sendfile() depends on the TCP_CORK socket option to avoid undesirable packet boundaries
• FreeBSD has a similar option called TCP_NOPUSH
• when TCP_CORK is turned off any buffered data is sent immediately, but this is not the case for TCP_NOPUSH

By default NGINX disable the use of TCP_NOPUSH option:

# http, server, location contexts
# To turn on tcp_nopush (my recommendation):
tcp_nopush on;

# To turn off tcp_nopush:
tcp_nopush off;   # default

Mixing all together

There are many opinions on this. My recommendation is to set all to on. However, I quote an interesting comment (Mixing sendfile, tcp_nodelay and tcp_nopush illogical?) that should dispel any doubts:

When set indicates to always queue non-full frames. Later the user clears this option and we transmit any pending partial frames in the queue. This is meant to be used alongside sendfile() to get properly filled frames when the user (for example) must write out headers with a write() call first and then use sendfile to send out the data parts. TCP_CORK can be set together with TCP_NODELAY and it is stronger than TCP_NODELAY.

Summarizing:

• tcp_nodelay on; is generaly at the odds with tcp_nopush on; as they are mutually exclusive
• NGINX has special behavior that if you have sendfile on;, it uses TCP_NOPUSH for everything but the last package
• and then turns TCP_NOPUSH off and enables TCP_NODELAY to avoid 200ms ACK delay

So in fact, the most important changes are listed below:

sendfile on;
tcp_nopush on;    # with this, the tcp_nodelay does not really matter

Request processing stages

When building filtering rules (e.g. with allow/deny) you should always remember to test them and to know what happens at each of the phases (which modules are used). For additional information about the potential problems, look at allow and deny section and Take care about your ACL rules — Hardening — P1.

There can be altogether 11 phases when NGINX handles (processes) a request:

• NGX_HTTP_POST_READ_PHASE — first phase, read the request header

  • example modules: ngx_http_realip_module

• NGX_HTTP_SERVER_REWRITE_PHASE — implementation of rewrite directives defined in a server block; to change request URI using PCRE regular expressions, return redirects, and conditionally select configurations

  • example modules: ngx_http_rewrite_module

• NGX_HTTP_FIND_CONFIG_PHASE — replace the location according to URI (location lookup)

• NGX_HTTP_REWRITE_PHASE — URI transformation on location level

  • example modules: ngx_http_rewrite_module

• NGX_HTTP_POST_REWRITE_PHASE — URI transformation post-processing (the request is redirected to a new location)

  • example modules: ngx_http_rewrite_module

• NGX_HTTP_PREACCESS_PHASE — authentication preprocessing request limit, connection limit (access restriction)

  • example modules: ngx_http_limit_req_module, ngx_http_limit_conn_module, ngx_http_realip_module

• NGX_HTTP_ACCESS_PHASE — verification of the client (the authentication process, limiting access)

  • example modules: ngx_http_access_module, ngx_http_auth_basic_module

• NGX_HTTP_POST_ACCESS_PHASE — access restrictions check post-processing phase, the certification process, processing satisfy any directive

  • example modules: ngx_http_access_module, ngx_http_auth_basic_module

• NGX_HTTP_PRECONTENT_PHASE — generating content

  • example modules: ngx_http_try_files_module

• NGX_HTTP_CONTENT_PHASE — content processing

  • example modules: ngx_http_index_module, ngx_http_autoindex_module, ngx_http_gzip_module

• NGX_HTTP_LOG_PHASE — log processing

  • example modules: ngx_http_log_module

You may feel lost now (me too…) so I let myself put this great and simple preview:

^{This infographic comes from Inside NGINX official library.}

On every phase you can register any number of your handlers. Each phase has a list of handlers associated with it.

I recommend to read a great explanation about HTTP request processing phases in Nginx and, of course, official Development guide. I have also prepared a simple diagram that can help you understand what modules are used in each phase. It also contains short descriptions from official development guide:

Server blocks logic

NGINX does have server blocks (like a virtual hosts in an Apache) that use listen directive to bind to TCP sockets and server_name directive to identify virtual hosts.

It’s a short example of two server block contexts with several regular expressions:

http {

  index index.html;
  root /var/www/example.com/default;

  server {

    listen 10.10.250.10:80;
    server_name www.example.com;

    access_log logs/example.access.log main;

    root /var/www/example.com/public;

    location ~ ^/(static|media)/ { ... }

    location ~* /[0-9][0-9](-.*)(.html)$ { ... }

    location ~* .(jpe?g|png|gif|ico)$ { ... }

    location ~* (?<begin>.*app)/(?<end>.+.php)$ { ... }

    ...

  }

  server {

    listen 10.10.250.11:80;
    server_name "~^(api.)?example.com api.de.example.com";

    access_log logs/example.access.log main;

    location ~ ^(/[^/]+)/api(.*)$ { ... }

    location ~ ^/backend/id/([a-z].[a-z]*) { ... }

    ...

  }

}

Handle incoming connections

🔖 Define the listen directives with address:port pair — Base Rules — P1
🔖 Prevent processing requests with undefined server names — Base Rules — P1
🔖 Never use a hostname in a listen or upstream directives — Base Rules — P1
🔖 Use exact names in a server_name directive if possible — Performance — P2
🔖 Separate listen directives for 80 and 443 ports — Base Rules — P3
🔖 Use only one SSL config for the listen directive — Base Rules — P3

NGINX uses the following logic to determining which virtual server (server block) should be used:

1. Match the address:port pair to the listen directive — that can be multiple server blocks with listen directives of the same specificity that can handle the request

   NGINX use the address:port combination for handle incoming connections. This pair is assigned to the listen directive.

   The listen directive can be set to:

   • an IP address/port combination (127.0.0.1:80;)

   • a lone IP address, if only address is given, the port 80 is used (127.0.0.1;) — becomes 127.0.0.1:80;

   • a lone port which will listen to every interface on that port (80; or *:80;) — becomes 0.0.0.0:80;

   • the path to a UNIX domain socket (unix:/var/run/nginx.sock;)

   If the listen directive is not present then either *:80 is used (runs with the superuser privileges), or *:8000 otherwise.

   To play with listen directive NGINX must follow the following steps:

   • NGINX translates all incomplete listen directives by substituting missing values with their default values (see above)

   • NGINX attempts to collect a list of the server blocks that match the request most specifically based on the address:port

   • If any block that is functionally using 0.0.0.0, will not be selected if there are matching blocks that list a specific IP

   • If there is only one most specific match, that server block will be used to serve the request

   • If there are multiple server blocks with the same level of matching, NGINX then begins to evaluate the server_name directive of each server block

   Look at this short example:

   # From client side:
   GET / HTTP/1.0
   Host: api.random.com
   
   # From server side:
   server {
   
     # This block will be processed:
     listen 192.168.252.10;  # --> 192.168.252.10:80
   
     ...
   
   }
   
   server {
   
     listen 80;  # --> *:80 --> 0.0.0.0:80
     server_name api.random.com;
   
     ...
   
   }

2. Match the Host header field against the server_name directive as a string (the exact names hash table)

3. Match the Host header field against the server_name directive with a
   wildcard at the beginning of the string (the hash table with wildcard names starting with an asterisk)

If one is found, that block will be used to serve the request. If multiple matches are found, the longest match will be used to serve the request.

1. Match the Host header field against the server_name directive with a
   wildcard at the end of the string (the hash table with wildcard names ending with an asterisk)

If one is found, that block is used to serve the request. If multiple matches are found, the longest match will be used to serve the request.

1. Match the Host header field against the server_name directive as a regular expression

The first server_name with a regular expression that matches the Host header will be used to serve the request.

1. If all the Host headers doesn’t match, then direct to the listen directive marked as default_server (makes the server block answer all the requests that doesn’t match any server block)

2. If all the Host headers doesn’t match and there is no default_server,
   direct to the first server with a listen directive that satisfies first step

3. Finally, NGINX goes to the location context

^{This list is based on Mastering Nginx — The virtual server section.}

Matching location

🔖 Make an exact location match to speed up the selection process — Performance — P3

For each request, NGINX goes through a process to choose the best location block that will be used to serve that request.

The location block enables you to handle several types of URIs/routes (Layer 7 routing based on URL), within a server block. Syntax looks like:

location optional_modifier location_match { ... }

location_match in the above defines what NGINX should check the request URI against. The optional_modifier below will cause the associated location block to be interpreted as follows (the order doesn’t matter at this moment):

• (none): if no modifiers are present, the location is interpreted as a prefix match. To determine a match, the location will now be matched against the beginning of the URI

• =: is an exact match, without any wildcards, prefix matching or regular expressions; forces a literal match between the request URI and the location parameter

• ~: if a tilde modifier is present, this location must be used for case sensitive matching (RE match)

• ~*: if a tilde and asterisk modifier is used, the location must be used for case insensitive matching (RE match)

• ^~: assuming this block is the best non-RE match, a carat followed by a tilde modifier means that RE matching will not take place

And now, a short introduction to determines location priority:

• the exact match is the best priority (processed first); ends search if match

• the prefix match is the second priority; there are two types of prefixes: ^~ and (none), if this match used the ^~ prefix, searching stops

• the regular expression match has the lowest priority; there are two types of prefixes: ~ and ~*; in the order they are defined in the configuration file

• if regular expression searching yielded a match, that result is used, otherwise, the match from prefix searching is used

So, look at this example, it comes from the Nginx documentation — ngx_http_core_module:

location = / {
  # Matches the query / only.
  [ configuration A ]
}
location / {
  # Matches any query, since all queries begin with /, but regular
  # expressions and any longer conventional blocks will be
  # matched first.
  [ configuration B ]
}
location /documents/ {
  # Matches any query beginning with /documents/ and continues searching,
  # so regular expressions will be checked. This will be matched only if
  # regular expressions don't find a match.
  [ configuration C ]
}
location ^~ /images/ {
  # Matches any query beginning with /images/ and halts searching,
  # so regular expressions will not be checked.
  [ configuration D ]
}
location ~* .(gif|jpg|jpeg)$ {
  # Matches any request ending in gif, jpg, or jpeg. However, all
  # requests to the /images/ directory will be handled by
  # Configuration D.
  [ configuration E ]
}

To help you understand how does location match works:

• Nginx location match tester
• Nginx location match visible
• NGINX Regular Expression Tester

The process of choosing NGINX location block is as follows (a detailed explanation):

1. NGINX searches for an exact match. If a = modifier (e.g. location = foo { ... }) exactly matches the request URI, this specific location block is chosen right away

• this block is processed
• match-searching stops

1. Prefix-based NGINX location matches (no regular expression). Each location will be checked against the request URI. If no exact (meaning no = modifier) location block is found, NGINX will continue with non-exact prefixes. It starts with the longest matching prefix location for this URI, with the following approach:

• In case the longest matching prefix location has the ^~ modifier (e.g. location ^~ foo { ... }), NGINX will stop its search right away and choose this location

  • the block of the longest (most explicit) of those matches is processed
  • match-searching stops

• Assuming the longest matching prefix location doesn’t use the ^~ modifier, the match is temporarily stored and the process continues

I’m not sure about the order. In the official documentation it is not clearly indicated and external guides explain it differently. It seems logical to check the longest matching prefix location first.

1. As soon as the longest matching prefix location is chosen and stored, NGINX continues to evaluate the case-sensitive (e.g. location ~ foo { ... }) and insensitive regular expression (e.g. location ~* foo { ... }) locations. The first regular expression location that fits the URI is selected right away to process the request

• the block of the first matching regex found (when parsing the config-file top-to-bottom) is processed
• match-searching stops

1. If no regular expression locations are found that match the request URI, the previously stored prefix location (e.g. location foo { ... }) is selected to serve the request

• location / kind of a catch all location
• the block of the longest (most explicit) of those matches is processed
• match-searching stops

You should also know, that the non-regex match-types are fully declarative — order of definition in the config doesn’t matter — but the winning regex-match (if processing even gets that far) is entirely based on its order of entry in the config file.

In order, to better understand how this process work, please see this short cheatsheet that will allow you to design your location blocks in a predictable way:

I recommend to use external tools for testing regular expressions. For more please see online tools chapter.

Ok, so here’s a more complicated configuration:

server {

 listen 80;
 server_name xyz.com www.xyz.com;

 location ~ ^/(media|static)/ {
  root /var/www/xyz.com/static;
  expires 10d;
 }

 location ~* ^/(media2|static2) {
  root /var/www/xyz.com/static2;
  expires 20d;
 }

 location /static3 {
  root /var/www/xyz.com/static3;
 }

 location ^~ /static4 {
  root /var/www/xyz.com/static4;
 }

 location = /api {
  proxy_pass http://127.0.0.1:8080;
 }

 location / {
  proxy_pass http://127.0.0.1:8080;
 }

 location /backend {
  proxy_pass http://127.0.0.1:8080;
 }

 location ~ logo.xcf$ {
  root /var/www/logo;
  expires 48h;
 }

 location ~* .(png|ico|gif|xcf)$ {
  root /var/www/img;
  expires 24h;
 }

 location ~ logo.ico$ {
  root /var/www/logo;
  expires 96h;
 }

 location ~ logo.jpg$ {
  root /var/www/logo;
  expires 48h;
 }

}

And look the table with the results:

rewrite vs return

Generally there are two ways of implementing redirects in NGINX: with rewrite and return directives.

These directives (comes from the ngx_http_rewrite_module) are very useful but (from the NGINX documentation) the only 100% safe things which may be done inside if in a location context are:

• return ...;
• rewrite ... last;

Anything else may possibly cause unpredictable behaviour, including potential SIGSEGV.

rewrite directive

The rewrite directives are executed sequentially in order of their appearance in the configuration file. It’s slower (but still extremely fast) than a return and returns HTTP 302 in all cases, irrespective of permanent.

The rewrite directive just changes the request URI, not the response of request. Importantly only the part of the original url that matches the regex is rewritten. It can be used for temporary url changes.

I sometimes used rewrite to capture elementes in the original URL, change or add elements in the path, and in general when I do something more complex:

location / {

  ...

  rewrite ^/users/(.*)$ /user.php?username=$1 last;

  # or:
  rewrite ^/users/(.*)/items$ /user.php?username=$1&page=items last;

}

You must know that rewrite returns only code 301 or 302.

rewrite directive accept optional flags:

• break — basically completes processing of rewrite directives, stops processing, and breakes location lookup cycle by not doing any location lookup and internal jump at all

  • if you use break flag inside location block:

    • no more parsing of rewrite conditions
    • internal engine continues to parse the current location block

    Inside a location block, with break, NGINX only stops processing anymore rewrite conditions.

  • if you use break flag outside location block:

    • no more parsing of rewrite conditions
    • internal engine goes to the next phase (searching for location match)

    Outside a location block, with break, NGINX stops processing anymore rewrite conditions.

• last — basically completes processing of rewrite directives, stops processing, and starts a search for a new location matching the changed URI

  • if you use last flag inside location block:

    • no more parsing of rewrite conditions
    • internal engine starts to look for another location match based on the result of the rewrite result
    • no more parsing of rewrite conditions, even on the next location match

    Inside a location block, with last, NGINX stops processing anymore rewrite conditions and then starts to look for a new matching of location block. NGINX also ignores any rewrites in the new location block.

  • if you use last flag outside location block:

    • no more parsing of rewrite conditions
    • internal engine goes to the next phase (searching for location match)

    Outside a location block, with last, NGINX stops processing anymore rewrite conditions.

• redirect — returns a temporary redirect with the 302 HTTP response code

• permanent — returns a permanent redirect with the 301 HTTP response code

Note:

• that outside location blocks, last and break are effectively the same
• processing of rewrite directives at server level may be stopped via break, but the location lookup will follow anyway

^{This explanation is based on the awesome answer by Pothi Kalimuthu to nginx url rewriting: difference between break and last.}

Official documentation has a great tutorials about Creating NGINX Rewrite Rules and Converting rewrite rules. I also recommend Clean Url Rewrites Using Nginx.

Finally, look at the difference between last and break flags in action:

• last directive:

• break directive:

^{This infographic comes from Internal rewrite — nginx by Ivan Dabic.}

return directive

🔖 Use return directive for URL redirection (301, 302) — Base Rules — P2
🔖 Use return directive instead of rewrite for redirects — Performance — P2

The other way is a return directive. It’s faster than rewrite because there is no regexp that has to be evaluated. It’s stops processing and returns HTTP 301 (by default) to a client (tells NGINX to respond directly to the request), and the entire url is rerouted to the url specified.

I use return directive in the following cases:

• force redirect from http to https:

  server {
  
    ...
  
    return 301 https://example.com$request_uri;
  
  }

• redirect from www to non-www and vice versa:

  server {
  
    ...
  
    # It's only example. You shouldn't use 'if' statement in the following case:
    if ($host = www.example.com) {
  
      return 301 https://example.com$request_uri;
  
    }
  
  }

• close the connection and log it internally:

  server {
  
    ...
  
    return 444;
  
  }

• send 4xx HTTP response for a client without any other actions:

  server {
  
    ...
  
    if ($request_method = POST) {
  
      return 405;
  
    }
  
    # or:
    if ($invalid_referer) {
  
      return 403;
  
    }
  
    # or:
    if ($request_uri ~ "^/app/(.+)$") {
  
      return 403;
  
    }
  
    # or:
    location ~ ^/(data|storage) {
  
      return 403;
  
    }
  
  }

• and sometimes for reply with HTTP code without serving a file or response body:

  server {
  
    ...
  
    # NGINX will not allow a 200 with no response body (200's need to be with a resource in the response.
    # '204 No Content' is meant to say "I've completed the request, but there is no body to return"):
    return 204 "it's all okay";
    # Or without body:
    return 204;
  
    # Because default Content-Type is application/octet-stream, browser will offer to "save the file".
    # If you want to see reply in browser you should add properly Content-Type:
    # add_header Content-Type text/plain;
  
  }

To the last example: be careful if you’re using such a configuration to do a healthcheck. While a 204 HTTP code is semantically perfect for a healthcheck (success indication with no content), some services do not consider it a success.

URL redirections

🔖 Use return directive for URL redirection (301, 302) — Base Rules — P2
🔖 Use return directive instead of rewrite for redirects — Performance — P2

HTTP allows servers to redirect a client request to a different location. This is useful when moving content to a new URL, when deleting pages or when changing domain names or merging websites.

URL redirection is done for various reasons:

• for URL shortening
• to prevent broken links when web pages are moved
• to allow multiple domain names belonging to the same owner to refer to a single web site
• to guide navigation into and out of a website
• for privacy protection
• for hostile purposes such as phishing attacks or malware distribution

^{It comes from Wikipedia — URL redirection.}

I recommend to read:

• Redirections in HTTP
• 301 101: How Redirects Work
• Modify 301/302 response body (from this handbook)
• Redirect POST request with payload to external endpoint (from this handbook)

try_files directive

We have one more very interesting and important directive: try_files (from the ngx_http_core_module). This directive tells NGINX to check for the existence of a named set of files or directories (checks files conditionally breaking on success).

I think the best explanation comes from the official documentation:

try_files checks the existence of files in the specified order and uses the first found file for request processing; the processing is performed in the current context. The path to a file is constructed from the file parameter according to the root and alias directives. It is possible to check directory’s existence by specifying a slash at the end of a name, e.g. $uri/. If none of the files were found, an internal redirect to the uri specified in the last parameter is made.

Generally it may check files on disk, redirect to proxies or internal locations, and return error codes, all in one directive.

Take a look at the following example:

server {

  ...

  root /var/www/example.com;

  location / {

    try_files $uri $uri/ /frontend/index.html;

  }

  location ^~ /images {

    root /var/www/static;
    try_files $uri $uri/ =404;

  }

  ...

• default root directory for all locations is /var/www/example.com

• location / — matches all locations without more specific locations, e.g. exact names

  • try_files $uri — when you receive a URI that’s matched by this block try $uri first

    For example: https://example.com/tools/en.js — NGINX will try to check if there’s a file inside /tools called en.js, if found it, serve it in the first place.

  • try_files $uri $uri/ — if you didn’t find the first condition try the URI as a directory

    For example: https://example.com/backend/ — NGINX will try first check if a file called backend exists, if can’t find it then goes to second check $uri/ and see if there’s a directory called backend exists then it will try serving it.

  • try_files $uri $uri/ /frontend/index.html — if a file and directory not found, NGINX sends /frontend/index.html

• location ^~ /images — handle any query beginning with /images and halts searching

  • default root directory for this location is /var/www/static

  • try_files $uri — when you receive a URI that’s matched by this block try $uri first

    For example: https://example.com/images/01.gif — NGINX will try to check if there’s a file inside /images called 01.gif, if found it, serve it in the first place.

  • try_files $uri $uri/ — if you didn’t find the first condition try the URI as a directory

    For example: https://example.com/images/ — NGINX will try first check if a file called images exists, if can’t find it then goes to second check $uri/ and see if there’s a directory called images exists then it will try serving it.

  • try_files $uri $uri/ =404 — if a file and directory not found, NGINX sends HTTP 404 (Not Found)

On the other hand, try_files is relatively primitive. When encountered, NGINX will look for any of the specified files physically in the directory matched by the location block. If they don’t exist, NGINX does an internal redirect to the last entry in the directive.

Additionally, think about dont’t check for the existence of directories:

# Use this to take out an extra filesystem stat():
try_files $uri @index;

# Instead of this:
try_files $uri $uri/ @index;

if, break and set

🔖 Avoid checks server_name with if directive — Performance — P2

The ngx_http_rewrite_module also provides additional directives:

• break — stops processing, if is specified inside the location, further processing of the request continues in this location:

  # It's useful for:
  if ($slow_resp) {
  
    limit_rate 50k;
    break;
  
  }

• if — you can use if inside a server but not the other way around, also notice that you shouldn’t use if inside location as it may not work as desired. For example, if statements aren’t a good way of setting custom headers because they may cause statements outside the if block to be ignored. The NGINX docs says:

  There are cases where you simply cannot avoid using an if, for example if you need to test a variable which has no equivalent directive.

  You should also remember about this:

  The if context in NGINX is provided by the rewrite module and this is the primary intended use of this context. Since NGINX will test conditions of a request with many other purpose-made directives, if should not be used for most forms of conditional execution. This is such an important note that the NGINX community has created a page called if is evil (yes, it’s really evil and in most cases not needed).

  A long time ago I found this:

  That’s actually not true and shows you don’t understand the problem with it. When the if statement ends with return directive, there is no problem and it’s safe to use.

  On the other hand, official documentation say:

  Directive if has problems when used in location context, in some cases it doesn’t do what you expect but something completely different instead. In some cases it even segfaults. It’s generally a good idea to avoid it if possible.

• set — sets a value for the specified variable. The value can contain text, variables, and their combination

Example of usage if and set directives:

# It comes from: https://gist.github.com/jrom/1760790:
if ($request_uri = /) {

  set $test A;

}

if ($host ~* example.com) {

  set $test "${test}B";

}

if ($http_cookie !~* "auth_token") {

  set $test "${test}C";

}

if ($test = ABC) {

  proxy_pass http://cms.example.com;
  break;

}

root vs alias

Placing a root or alias directive in a location block overrides the root or alias directive that was applied at a higher scope.

With alias you can map to another file name. With root forces you to name your file on the server. In the first case, NGINX replaces the string prefix e.g /robots.txt in the URL path with e.g. /var/www/static/robots.01.txt and then uses the result as a filesystem path. In the second, NGINX inserts the string e.g. /var/www/static/ at the beginning of the URL path and then uses the result as a file system path.

Look at this. There is a difference, when the alias is for a whole directory will work:

location ^~ /data/ { alias /home/www/static/data/; }

But the following code won’t do:

location ^~ /data/ { root /home/www/static/data/; }

This would have to be:

location ^~ /data/ { root /home/www/static/; }

The root directive is typically placed in server and location blocks. Placing a root directive in the server block makes the root directive available to all location blocks within the same server block.

This directive tells NGINX to take the request url and append it behind the specified directory. For example, with the following configuration block:

server {

  server_name example.com;
  listen 10.250.250.10:80;

  index index.html;
  root /var/www/example.com;

  location / {

    try_files $uri $uri/ =404;

  }

  location ^~ /images {

    root /var/www/static;
    try_files $uri $uri/ =404;

  }

}

NGINX will map the request made to:

• http://example.com/images/logo.png into the file path /var/www/static/images/logo.png
• http://example.com/contact.html into the file path /var/www/example.com/contact.html
• http://example.com/about/us.html into the file path /var/www/example.com/about/us.html

Like you want to forward all requests which start /static and your data present in /var/www/static you should set:

• first path: /var/www
• last path: /static
• full path: /var/www/static

location <last path> {

  root <first path>;

  ...

}

NGINX documentation on the alias directive suggests that it is better to use root over alias when the location matches the last part of the directive’s value.

The alias directive can only be placed in a location block. The following is a set of configurations for illustrating how the alias directive is applied:

server {

  server_name example.com;
  listen 10.250.250.10:80;

  index index.html;
  root /var/www/example.com;

  location / {

    try_files $uri $uri/ =404;

  }

  location ^~ /images {

    alias /var/www/static;
    try_files $uri $uri/ =404;

  }

}

NGINX will map the request made to:

• http://example.com/images/logo.png into the file path /var/www/static/logo.png
• http://example.com/images/ext/img.png into the file path /var/www/static/ext/img.png
• http://example.com/contact.html into the file path /var/www/example.com/contact.html
• http://example.com/about/us.html into the file path /var/www/example.com/about/us.html

When location matches the last part of the directive’s value it is better to use the root directive (it seems like an arbitrary style choice because authors don’t justify that instruction at all). Look at this example from the official documentation:

location /images/ {

  alias /data/w3/images/;

}

# Better solution:
location /images/ {

  root /data/w3;

}

internal directive

This directive specifies that the location block is internal. In other words,
the specified resource cannot be accessed by external requests.

On the other hand, it specifies how external redirections, i.e. locations like http://example.com/app.php/some-path should be handled; while set, they should return 404, only allowing internal redirections. In brief, this tells NGINX it’s not accessible from the outside (it doesn’t redirect anything).

Conditions handled as internal redirections are listed in the documentation for internal directive. Specifies that a given location can only be used for internal requests and are the following:

• requests redirected by the error_page, index, random_index, and try_files directives
• requests redirected by the X-Accel-Redirect response header field from an upstream server
• subrequests formed by the include virtual command of the ngx_http_ssi_module module, by the ngx_http_addition_module module directives, and by auth_request and mirror directives
• requests changed by the rewrite directive

Example 1:

error_page 404 /404.html;

location = /404.html {

  internal;

}

Example 2:

The files are served from the directory /srv/hidden-files by the path prefix /hidden-files/. Pretty straightforward. The internal declaration tells NGINX that this path is accessible only through rewrites in the NGINX config, or via the X-Accel-Redirect header in proxied responses.

To use this, just return an empty response which contains that header. The content of the header should be the location you want to redirect to:

location /hidden-files/ {

  internal;
  alias /srv/hidden-files/;

}

Example 3:

Another use case for internal redirects in NGINX is to hide credentials. Often you need to make requests to 3rd party services. For example, you want to send text messages or access a paid maps server. It would be the most efficient to send these requests directly from your JavaScript front end. However, doing so means you would have to embed an access token in the front end. This means savvy users could extract this token and make requests on your account.

An easy fix is to make an endpoint in your back end which initiates the actual request. We could make use of an HTTP client library inside the back end. However, this will again tie up workers, especially if you expect a barrage of requests and the 3rd party service is responding very slowly.

location /external-api/ {

  internal;
  set $redirect_uri "$upstream_http_redirect_uri";
  set $authorization "$upstream_http_authorization";

  # For performance:
  proxy_buffering off;
  # Pass on secret from backend:
  proxy_set_header Authorization $authorization;
  # Use URI determined by backend:
  proxy_pass $redirect_uri;

}

^{Examples 2 and 3 (both are great!) comes from How to use internal redirects in NGINX.}

There is a limit of 10 internal redirects per request to prevent request processing cycles that can occur in incorrect configurations. If this limit is reached, the error HTTP 500 Internal Server Error is returned. In such cases, the rewrite or internal redirection cycle message can be seen in the error log.

Look also at Authentication Based on Subrequest Result from the official documentation.

External and internal redirects

External redirects originate directly from the client. So, if the client fetched https://example.com/directory it would be directly fall into preceding location block.

Internal redirect means that it doesn’t send a 302 response to the client, it simply performs an implicit rewrite of the url and attempts to process it as though the user typed the new url originally.

The internal redirect is different from the external redirect defined by HTTP response code 302 and 301, client browser won’t update its URI addresses.

To begin rewriting internally, we should explain the difference between redirects and internal rewrite. When source points to a destination that is out of source domain that is what we call redirect as your request will go from source to outside domain/destination.

With internal rewrite you would be, basically, doing the same only the destination is local path under same domain and not the outside location.

There is also great explanation about internal redirects:

The internal redirection (e.g. via the echo_exec or rewrite directive) is an operation that makes NGINX jump from one location to another while processing a request (are very similar to goto statement in the C language). This «jumping» happens completely within the server itself.

There are two different kinds of internal requests:

• internal redirects — redirects the client requests internally. The URI is
  changed, and the request may therefore match another location block and
  become eligible for different settings. The most common case of internal
  redirects is when using the rewrite directive, which allows you to rewrite the
  request URI

• sub-requests — additional requests that are triggered internally to generate (insert or append to the body of the original request) content that is complementary to the main request (addition or ssi modules)

allow and deny

🔖 Take care about your ACL rules — Hardening — P1
🔖 Reject unsafe HTTP methods — Hardening — P1

Both comes from the ngx_http_access_module module and allows limiting access to certain client addresses. You can combining allow/deny rules.

deny will always return 403 error code.

The easiest path would be to start out by denying all access, then only granting access to those locations you want. For example:

location / {

  # without 'satisfy any' both should be passed:
  satisfy any;
  allow 192.168.0/0/16;
  deny all;

  # sh -c "echo -n 'user:' >> /etc/nginx/.secret"
  # sh -c "openssl passwd -apr1 >> /etc/nginx/.secret"
  auth_basic "Restricted Area";
  auth_basic_user_file /etc/nginx/.secret;

  root   /usr/share/nginx/html;
  index  index.html index.htm;

}

Putting satisfy any; in your configuration tells NGINX to accept either http authentication, or IP restriction. By default, when you define both, it will expect both.

See also this answer:

As you’ve found, it isn’t advisable to but the auth settings at the server level because they will apply to all locations. While it is possible to turn basic auth off there doesn’t appear to be a way to clear an existing IP whitelist.

A better solution would be to add the authentication to the / location so that it isn’t inherited by /hello.

The problem comes if you have other locations that require the basic auth and IP whitelisting in which case it might be worth considering moving the auth components to an include file or nesting them under /.

Both directives may work unexpectedly! Look at the following example:

server {

  server_name example.com;

  deny all;

  location = /test {
    return 200 "it's all okay";
    more_set_headers 'Content-Type: text/plain';
  }

}

If you generate a reqeust:

curl -i https://example.com/test
HTTP/2 200
date: Wed, 11 Nov 2018 10:02:45 GMT
content-length: 13
server: Unknown
content-type: text/plain

it's all okay

Why? Look at Request processing stages chapter. That’s because NGINX process request in phases, and rewrite phase (where return belongs) goes before access phase (where deny works).

uri vs request_uri

🔖 Use $request_uri to avoid using regular expressions — Performance — P2

$request_uri is the original request (for example /foo/bar.php?arg=baz includes arguments and can’t be modified) but $uri refers to the altered URI so $uri is not equivalent to $request_uri.

See this great and short explanation by Richard Smith:

The $uri variable is set to the URI that NGINX is currently processing — but it is also subject to normalisation, including:

• removal of the ? and query string
• consecutive / characters are replace by a single /
• URL encoded characters are decoded

The value of $request_uri is always the original URI and is not subject to any of the above normalisations.

Most of the time you would use $uri, because it is normalised. Using $request_uri in the wrong place can cause URL encoded characters to become doubly encoded.

Both excludes the schema (https:// and the port (implicit 443) in both examples above) as defined by RFC 2616 — http URL ^[IETF] for the URL:

http_URL = "http(s):" "//" host [ ":" port ] [ abs_path [ "?" query ]]

Take a look at the following table:

Another way to repeat the location is to use the proxy_pass directive which is quite easy:

location /app/ {

  proxy_pass http://127.0.0.1:5000;

  # or:
  proxy_pass http://127.0.0.1:5000/api/app/;

}

Compression and decompression

🔖 Mitigation of CRIME/BREACH attacks — Hardening Rules — P2

By default, NGINX compresses responses only with MIME type text/html using the gzip method. So, if you send request with Accept-Encoding: gzip header you will not see the Content-Encoding: gzip in the response.

To enable gzip compression:

To compress responses with other MIME types, include the gzip_types directive and list the additional types:

gzip_types text/plain text/css text/xml text/javascript application/x-javascript application/xml;

Remember: by default, NGINX doesn’t compress image files using its per-request gzip module.

I also highly recommend you read this (it’s interesting observation about gzip and performance by Barry Pollard):

To be honest gzip is not very processor intensive these days and gzipping on the fly (and then unzipping in the browser) is often the norm. It’s something web browsers are very good at.

So unless you are getting huge volumes of traffic you’ll probably not notice any performance or CPU load impact due to on the fly gzipping for most web files.

To test HTTP and Gzip compression I recommend two external tools:

• HTTP Compression Test
• HTTP Gzip Compression Test

NGINX also compress large files and avoid the temptation to compress smaller files (such as images, executables, etc.), because very small files barely benefit from compression. You can tell NGINX not to compress files smaller than e.g. 128 bytes:

For more information see Finding the Nginx gzip_comp_level Sweet Spot.

Compressing resources on-the-fly adds CPU-load and latency (wait for the compression to be done) every time a resource is served. NGINX also provides static compression with static module. It is better, for 2 reasons:

• you don’t have to gzip for each request
• you can use a higher gzip level

For example:

# Enable static gzip compression:
location ^~ /assets/ {

  gzip_static on;

  ...

}

You should put the gzip_static on; inside the blocks that configure static files, but if you’re only running one site, it’s safe to just put it in the http block.

NGINX does not automatically compress the files for you. You will have to do this yourself.

To compress files manually:

cd assets/
while IFS='' read -r -d '' _fd; do

  gzip -N4c ${_fd} > ${_fd}.gz

done < <(find . -maxdepth 1 -type f -regex ".*.(css|js|jpg|gif|png|jpeg)" -print0)

So, for example, to service a request for /foo/bar/file, NGINX tries to find and send the file /foo/bar/file.gz that directly, so no extra CPU-cost or latency is added to your requests, speeding up the serving of your app.

What is the best NGINX compression gzip level?

The level of gzip compression simply determines how compressed the data is on a scale from 1-9, where 9 is the most compressed. The trade-off is that the most compressed data usually requires the most work to compress/decompress but look also at this great answer. Author explains that the level of gzip compression doesn’t affect the difficulty to decompress.

I think the ideal compression level seems to be between 4 and 6. The following directive set how much files will be compressed:

Hash tables

Before start reading this chapter I recommend Hash tables explained.

To assist with the rapid processing of requests, NGINX uses hash tables. NGINX hash, though in principle is same as typical hash lists, but it has significant differences.

They are not meant for applications that add and remove elements dynamicall but are specifically designed to hold set of init time elements arranged in hash list. All elements that are put in the hash list are known while creating the hash list itself. No dynamic addtion or deletion is possible here.

This hash table is constructed and compiled during restart or reload and afterwards it’s running very fast. Main purpose seems to be speeding up the lookup of one time added elements.

Look at the Setting up hashes from official documentation:

To quickly process static sets of data such as server names, map directive’s values, MIME types, names of request header strings, NGINX uses hash tables. During the start and each re-configuration NGINX selects the minimum possible sizes of hash tables such that the bucket size that stores keys with identical hash values does not exceed the configured parameter (hash bucket size). The size of a table is expressed in buckets. The adjustment is continued until the table size exceeds the hash max size parameter. Most hashes have the corresponding directives that allow changing these parameters.

I also recommend Optimizations section and nginx — Hashing scheme explanation.

Some important information (based on this amazing research by brablc):

• the general recommendation would be to keep both values as small as possible and as less collisions as possible (during startup and with each reconfiguration, NGINX selects the smallest possible size for the hash tables)

• it depends on your setup, you can reduce the number of server from the table and reload the NGINX instead of restart

• if NGINX gave out communication about the need for increasing hash_max_size or hash_bucket_size, then it is first necessary to increase the first parameter

• bigger hash_max_size uses more memory, bigger hash_bucket_size uses more CPU cycles during lookup and more transfers from main memory to cache. If you have enough memory increase hash_max_size and try to keep hash_bucket_size as low as possible

• each hash table entry consumes space in a bucket. The space required is the length of the key (with some overhead to store the domain’s actual length as well), e.g. domain name

  Since stage.api.example.com is 21 characters, all entries consume at least 24 bytes in a bucket, and most consume 32 bytes or more.

• as you increase the number of entries, you have to increase the size of the hash table and/or the number of hash buckets in the table

  If NGINX complains increase hash_max_size first as long as it complains. If the number exceeds some big number (32769 for instance), increase hash_bucket_size to multiple of default value on your platform as long as it complains. If it does not complain anymore, decrease hash_max_size back as long as it does not complain. Now you have the best setup for your set of server names (each set of server names may need different setup).

• with a hash bucket size of 64 or 128, a bucket is full after 4 or 5 entries hash to it

• hash_max_size is not related to number of server names directly, if number of servers doubles, you may need to increase hash_max_size 10 times or even more to avoid collisions. If you cannot avoid them, you have to increase hash_bucket_size

• if you have hash_max_size less than 10000 and small hash_bucket_size, you can expect long loading time because NGINX would try to find optimal hash size in a loop (see src/core/ngx_hash.c)

• if you have hash_max_size bigger than 10000, there will be only 1000 loops performed before it would complain

Server names hash table

The hash with the names of servers are controlled by the following directives (inside http context):

• server_names_hash_max_size — sets the maximum size of the server names hash tables; default value: 512

• server_names_hash_bucket_size — sets the bucket size for the server names hash tables; default values: 32, 64, or 128 (the default value depends on the size of the processor’s cache line)

  Parameter server_names_hash_bucket_size is always equalized to the size, multiple to the size of the line of processor cache.

If server name is defined as too.long.server.name.example.com then NGINX will fail to start and display the error message like:

nginx: [emerg] could not build server_names_hash, you should increase server_names_hash_bucket_size: 64

To fix this, you should reload the NGINX or increase the server_names_hash_bucket_size directive value to the next power of two (in this case to 128).

If a large number of server names are defined, and NGINX complained with the following error:

nginx: [emerg] could not build the server_names_hash, you should increase either server_names_hash_max_size: 512 or server_names_hash_bucket_size: 32

Try to set the server_names_hash_max_size to a number close to the number of server names. Only if this does not help, or if NGINX’s start time is unacceptably long, try to increase the server_names_hash_bucket_size parameter.

Log files

🔖 Use custom log formats — Debugging — P4

Log files are a critical part of the NGINX management. It writes information about client requests in the access log right after the request is processed (in the last phase: NGX_HTTP_LOG_PHASE).

By default:

• the access log is located in logs/access.log, but I suggest you take it to /var/log/nginx directory
• data is written in the predefined combined/main format
• access.log stores record of each request and log format is fully configurable
• error.log contains important operational messages

It is the equivalent to the following configuration:

# In nginx.conf (default log format):
http {

  ...

  log_format main
                  '$remote_addr - $remote_user [$time_local] "$request" '
                  '$status $body_bytes_sent "$http_referer" '
                  '"$http_user_agent" "$http_x_forwarded_for"';

  # but I suggest you change:
  log_format main
                  '$remote_addr - $remote_user [$time_local] '
                  '"$request_method $scheme://$host$request_uri '
                  '$server_protocol" $status $body_bytes_sent '
                  '"$http_referer" "$http_user_agent" '
                  '$request_time';

}

For more information please see Configuring Logging.

Set access log off; to completely turns off logging.

If you don’t want 404 errors to show in your NGINX error logs, you should set log_not_found off;.

If you want to enable logging of subrequests into access_log, you should set log_subrequest on; and change the default logging format (you have to log $uri to see the difference). There is great explanation about how to identify subrequests in NGINX log files.

I also recommend to read:

• ngx_http_log_module
• ngx_http_upstream_module

Conditional logging

Sometimes certain entries are there just to fill up the logs or are cluttering them. I sometimes exclude requests — by client IP or whatever else — when I want to debug log files more effective.

So, in this example, if the $error_codes variable’s value is 0 — then log nothing (default action), but if 1 (e.g. 404 or 503 from backend) — to save this request to the log:

# Define map in the http context:
http {

  ...

  map $status $error_codes {

    default   1;
    ~^[23]    0;

  }

  ...

  # Add if condition to the access log:
  access_log /var/log/nginx/example.com-access.log combined if=$error_codes;

}

Manually log rotation

🔖 Configure log rotation policy — Base Rules — P1

NGINX will re-open its logs in response to the USR1 signal:

cd /var/log/nginx

mv access.log access.log.0
kill -USR1 $(cat /var/run/nginx.pid) && sleep 1

# >= gzip-1.6:
gzip -k access.log.0
# With any version:
gzip < access.log.0 > access.log.0.gz

# Test integrity and remove if test passed:
gzip -t access.log.0 && rm -fr access.log.0

Error log severity levels

You can’t specify your own format, but in NGINX build-in several level’s of error_log-ing.

The following is a list of all severity levels:

For example: if you set crit error log level, messages of crit, alert, and emerg levels are logged.

For debug logging to work, NGINX needs to be built with --with-debug.

Default values for the error level:

• in the main section — error
• in the HTTP section — crit
• in the server section — crit

How to log the start time of a request?

The most logging information requires the request to complete (status code, bytes sent, durations, etc). If you want to log the start time of a request in NGINX you should apply a patch that exposes request start time as a variable.

The $time_local variable contains the time when the log entry is written so when the HTTP request header is read, NGINX does a lookup of the associated virtual server configuration. If the virtual server is found, the request goes through six phases:

• server rewrite phase
• location phase
• location rewrite phase (which can bring the request back to the previous phase)
• access control phase
• try_files phase
• log phase

Since the log phase is the last one, $time_local variable is much more close to the end of the request than it’s start.

How to log the HTTP request body?

Nginx doesn’t parse the client request body unless it really needs to, so it usually does not fill the $request_body variable.

The exceptions are when:

• it sends the request to a proxy
• or a fastcgi server

So you really need to either add the proxy_pass or fastcgi_pass directives to your block.

# 1) Set log format:
log_format req_body_logging '$remote_addr - $remote_user [$time_local] '
                            '"$request" $status $body_bytes_sent '
                            '"$http_referer" "$http_user_agent" "$request_body"';

# 2) Limit the request body size:
client_max_body_size 1k;
client_body_buffer_size 1k;
client_body_in_single_buffer on;

# 3) Put the log format:
server {

  ...

  location /api/v4 {

    access_log logs/access_req_body.log req_body_logging;
    proxy_pass http://127.0.0.1;

    ...

  }

  location = /post.php {

    access_log /var/log/nginx/postdata.log req_body_logging;
    fastcgi_pass php_cgi;

    ...

  }

}

For this, you can also use echo module. To log a request body, what we need is to use the echo_read_request_body directive and the $request_body variable (contains the request body of the echo module).

echo_read_request_body explicitly reads request body so that the $request_body variable will always have non-empty values (unless the body is so big that it has been saved by NGINX to a local temporary file).

http {

  log_format req_body_logging '$request_body';
  access_log /var/log/nginx/access.log req_body_logging;

  ...

  server {

    location / {

      echo_read_request_body;

      ...

    }

    ...

  }

}

NGINX upstream variables returns 2 values

For example:

upstream_addr 192.168.50.201:8080 : 192.168.50.201:8080
upstream_bytes_received 427 : 341
upstream_connect_time 0.001 : 0.000
upstream_header_time 0.003 : 0.001
upstream_response_length 0 : 0
upstream_response_time 0.003 : 0.001
upstream_status 401 : 200

Below is a short description of each of them:

• $upstream_addr — keeps the IP address and port, or the path to the UNIX-domain socket of the upstream server. If several servers were contacted during request processing, their addresses are separated by commas, e.g. 192.168.1.1:80, 192.168.1.2:80, unix:/tmp/sock. If an internal redirect from one server group to another happens, initiated by X-Accel-Redirect or error_page, then the server addresses from different groups are separated by colons, e.g. 192.168.1.1:80, 192.168.1.2:80, unix:/tmp/sock : 192.168.10.1:80, 192.168.10.2:80
• $upstream_cache_status — keeps the status of accessing a response cache (0.8.3). The status can be either MISS, BYPASS, EXPIRED, STALE, UPDATING, REVALIDATED, or HIT
• $upstream_connect_time — time spent on establishing a connection with an upstream server
• $upstream_cookie_ — cookie with the specified name sent by the upstream server in the Set-Cookie response header field (1.7.1). Only the cookies from the response of the last server are saved
• $upstream_header_time — time between establishing a connection and receiving the first byte of the response header from the upstream server
• $upstream_http_ — keep server response header fields. For example, the Server response header field is available through the $upstream_http_server variable. The rules of converting header field names to variable names are the same as for the variables that start with the $http_ prefix. Only the header fields from the response of the last server are saved
• $upstream_response_length — keeps the length of the response obtained from the upstream server (0.7.27); the length is kept in bytes. Lengths of several responses are separated by commas and colons like addresses in the $upstream_addr variable
• $upstream_response_time — time between establishing a connection and receiving the last byte of the response body from the upstream server
• $upstream_status — keeps status code of the response obtained from the upstream server. Status codes of several responses are separated by commas and colons like addresses in the $upstream_addr variable

Official documentation say:

[…] If several servers were contacted during request processing, their addresses are separated by commas. […] If an internal redirect from one server group to another happens, initiated by “X-Accel-Redirect” or error_page, then the server addresses from different groups are separated by colons

This means that it made multiple requests to a backend, most likely you either have a bare proxy_pass host that resolves to different IPs (frequently the case with something like Amazon ELB as an origin), are you have a configured upstream that has multiple servers. Unless disabled, the proxy module will make round robin attempts against all healthy backends. This can be configured from proxy_next_upstream_* directives.

For example if this is not the desired behavior, you can just do (specifies in which cases a request should be passed to the next server):

# One should bear in mind that passing a request to the next server is only possible
# if nothing has been sent to a client yet. That is, if an error or timeout occurs
# in the middle of the transferring of a response, fixing this is impossible.
proxy_next_upstream off;

For more information please see ngx_http_upstream_module and proxy_next_upstream.

Reverse proxy

After reading this chapter, please see: Rules: Reverse Proxy.

This is one of the greatest feature of the NGINX. In simplest terms, a reverse proxy is a server that comes in-between internal applications and external clients, forwarding client requests to the appropriate server. It takes a client request, passes it on to one or more servers, and subsequently delivers the server’s response back to the client.

Official NGINX documentation says:

Proxying is typically used to distribute the load among several servers, seamlessly show content from different websites, or pass requests for processing to application servers over protocols other than HTTP.

You can also read a very good explanation about What’s the difference between proxy server and reverse proxy server.

A reverse proxy can off load much of the infrastructure concerns of a high-volume distributed web application.

^{This infographic comes from Jenkins with NGINX — Reverse proxy with https.}

This allow you to have NGINX reverse proxy requests to unicorns, mongrels, webricks, thins, or whatever you really want to have run your servers.

Reverse proxy gives you number of advanced features such as:

• load balancing, failover, and transparent maintenance of the backend servers
• increased security (e.g. SSL termination, hide upstream configuration)
• increased performance (e.g. caching, load balancing)
• simplifies the access control responsibilities (single point of access and maintenance)
• centralised logging and auditing (single point of maintenance)
• add/remove/modify HTTP headers

In my opinion, the two most important things related to the reverse proxy are:

• the way of requests forwarded to the backend
• the type of headers forwarded to the backend

If we talking about security of the proxy server look at this recommendations about Guidelines on Securing Public Web Servers ^[NIST]. This document is a good starting point. Is old but still has interesting solutions and suggestions.

There is a great explanation about the benefits of improving security through the use of a reverse proxy server.

A reverse proxy gives you a couple things that may make your server more secure:

• a place to monitor and log what is going on separate from the web server
• a place to filter separate from your web server if you know that some area of your system is vulnerable. Depending on the proxy you may be able to filter at the application level
• another place to implement ACLs and rules if you cannot be expressive enough for some reason on your web server
• a separate network stack that will not be vulnerable in the same ways as your web server. This is particularly true if your proxy is from a different vendor
• a reverse proxy with no filtering does not automatically protect you against everything, but if the system you need to protect is high-value then adding a reverse proxy may be worth the costs support and performance costs

Another great answer about best practices for reverse proxy implementation:

In my experience some of the most important requirements and mitigations, in no particular order, are:

• make sure that your proxy, back-end web (and DB) servers cannot establish direct outbound (internet) connections (including DNS and SMTP, and particularly HTTP). This means (forward) proxies/relays for required outbound access, if required
• make sure your logging is useful (§9.1 in the above), and coherent. You may have logs from multiple devices (router, firewall/IPS/WAF, proxy, web/app servers, DB servers). If you can’t quickly, reliably and deterministically link records across each device together, you’re doing it wrong. This means NTP, and logging any or all of: PIDs, TIDs, session-IDs, ports, headers, cookies, usernames, IP addresses and maybe more (and may mean some logs contain confidential information)
• understand the protocols, and make deliberate, informed decisions: including cipher/TLS version choice, HTTP header sizes, URL lengths, cookies. Limits should be implemented on the reverse-proxy. If you’re migrating to a tiered architecture, make sure the dev team are in the loop so that problems are caught as early as possible
• run vulnerability scans from the outside, or get someone to do it for you. Make sure you know your footprint and that the reports highlight deltas, as well as the theoretical TLS SNAFU du-jour
• understand the modes of failure. Sending users a bare default «HTTP 500 — the wheels came off» when you have load or stability problems is sloppy
• monitoring, metrics and graphs: having normal and historic data is invaluable when investigating anomalies, and for capacity planning
• tuning: from TCP time-wait to listen backlog to SYN-cookies, again you need to make make deliberate, informed decisions
• follow basic OS hardening guidelines, consider the use of chroot/jails, host-based IDS, and other measures, where available

Passing requests

🔖 Use pass directive compatible with backend protocol — Reverse Proxy — P1

When NGINX proxies a request, it sends the request to a specified proxied server, fetches the response, and sends it back to the client.

It is possible to proxy requests to:

• an HTTP servers (e.g. NGINX, Apache, or other) with proxy_pass directive:

  upstream bk_front {
  
    server 192.168.252.20:8080 weight=5;
    server 192.168.252.21:8080
  
  }
  
  server {
  
    location / {
  
      proxy_pass http://bk_front;
  
    }
  
    location /api {
  
      proxy_pass http://192.168.21.20:8080;
  
    }
  
    location /info {
  
      proxy_pass http://localhost:3000;
  
    }
  
    location /ra-client {
  
      proxy_pass http://10.0.11.12:8080/guacamole/;
  
    }
  
    location /foo/bar/ {
  
      proxy_pass http://www.example.com/url/;
  
    }
  
    ...
  
  }

• a non-HTTP servers (e.g. PHP, Node.js, Python, Java, or other) with proxy_pass directive (as a fallback) or directives specially designed for this:

  • fastcgi_pass which passes a request to a FastCGI server (PHP FastCGI Example):

    server {
    
      ...
    
      location ~ ^/.+.php(/|$) {
    
        fastcgi_pass 127.0.0.1:9000;
        include /etc/nginx/fcgi_params;
    
      }
    
      ...
    
    }

  • uwsgi_pass which passes a request to a uWSGI server (Nginx support uWSGI):

    server {
    
      location / {
    
        root html;
        uwsgi_pass django_cluster;
        uwsgi_param UWSGI_SCRIPT testapp;
        include /etc/nginx/uwsgi_params;
    
      }
    
      ...
    
    }

  • scgi_pass which passes a request to an SCGI server:

    server {
    
      location / {
    
        scgi_pass 127.0.0.1:4000;
        include /etc/nginx/scgi_params;
    
      }
    
      ...
    
    }

  • memcached_pass which passes a request to a Memcached server:

    server {
    
      location / {
    
        set $memcached_key "$uri?$args";
        memcached_pass memc_instance:4004;
    
        error_page 404 502 504 = @memc_fallback;
    
      }
    
      location @memc_fallback {
    
        proxy_pass http://backend;
    
      }
    
      ...
    
    }

  • redis_pass which passes a request to a Redis server (HTTP Redis):

    server {
    
      location / {
    
        set $redis_key $uri;
    
        redis_pass redis_instance:6379;
        default_type text/html;
        error_page 404 = /fallback;
    
      }
    
      location @fallback {
    
        proxy_pass http://backend;
    
      }
    
      ...
    
    }

The proxy_pass and other *_pass directives specifies that all requests which match the location block should be forwarded to the specific socket, where the backend app is running.

However, more complex apps may need additional directives:

• proxy_pass — see ngx_http_proxy_module directives explanation
• fastcgi_pass — see ngx_http_fastcgi_module directives explanation
• uwsgi_pass — see ngx_http_uwsgi_module directives explanation
• scgi_pass — see ngx_http_scgi_module directives explanation
• memcached_pass — see ngx_http_memcached_module directives explanation
• redis_pass — see ngx_http_redis_module directives explanation

Trailing slashes

🔖 Be careful with trailing slashes in proxy_pass directive — Reverse Proxy — P3

If you have something like:

location /public/ {

  proxy_pass http://bck_testing_01;

}

And go to http://example.com/public, NGINX will automatically redirect you to http://example.com/public/.

Look also at this example:

location /foo/bar/ {

  # proxy_pass http://example.com/url/;
  proxy_pass http://192.168.100.20/url/;

}

If the URI is specified along with the address, it replaces the part of the request URI that matches the location parameter. For example, here the request with the /foo/bar/page.html URI will be proxied to http://www.example.com/url/page.html.

If the address is specified without a URI, or it is not possible to determine the part of URI to be replaced, the full request URI is passed (possibly, modified).

Here is an example with trailing slash in location, but no trailig slash in proxy_pass:

location /foo/ {

  proxy_pass http://127.0.0.1:8080/bar;

}

See how bar and path concatenates. If one go to http://yourserver.com/foo/path/id?param=1 NGINX will proxy request to http://127.0.0.1/barpath/id?param=1.

As stated in NGINX documentation if proxy_pass used without URI (i.e. without path after server:port) NGINX will put URI from original request exactly as it was with all double slashes, ../ and so on.

Look also at the configuration snippets: Using trailing slashes.

Below are additional examples:

In other words:

You usually always want a trailing slash, never want to mix with and without trailing slash, and only want without trailing slash when you want to concatenate a certain path component together (which I guess is quite rarely the case). Note how query parameters are preserved.

Passing headers to the backend

🔖 Set the HTTP headers with add_header and proxy_*_header directives properly — Base Rules — P1
🔖 Remove support for legacy and risky HTTP headers — Hardening — P1
🔖 Always pass Host, X-Real-IP, and X-Forwarded headers to the backend — Reverse Proxy — P2
🔖 Use custom headers without X- prefix — Reverse Proxy — P3

By default, NGINX redefines two header fields in proxied requests:

• the Host header is re-written to the value defined by the $proxy_host variable. This will be the IP address or name and port number of the upstream, directly as defined by the proxy_pass directive

• the Connection header is changed to close. This header is used to signal information about the particular connection established between two parties. In this instance, NGINX sets this to close to indicate to the upstream server that this connection will be closed once the original request is responded to. The upstream should not expect this connection to be persistent

When NGINX proxies a request, it automatically makes some adjustments to the request headers it receives from the client:

• NGINX drop empty headers. There is no point of passing along empty values to another server; it would only serve to bloat the request

• NGINX, by default, will consider any header that contains underscores as invalid. It will remove these from the proxied request. If you wish to have NGINX interpret these as valid, you can set the underscores_in_headers directive to on, otherwise your headers will never make it to the backend server. Underscores in header fields are allowed (RFC 7230, sec. 3.2.), but indeed uncommon

It is important to pass more than just the URI if you expect the upstream server handle the request properly. The request coming from NGINX on behalf of a client will look different than a request coming directly from a client.

Please read Managing request headers from the official wiki.

In NGINX does support arbitrary request header field. Last part of a variable name is the field name converted to lower case with dashes replaced by underscores:

$http_name_of_the_header_key

If you have X-Real-IP = 127.0.0.1 in header, you can use $http_x_real_ip to get 127.0.0.1.

Use the proxy_set_header directive to sets headers that sends to the backend servers.

HTTP headers are used to transmit additional information between client and server. add_header sends headers to the client (browser) and will work on successful requests only, unless you set up always parameter. proxy_set_header sends headers to the backend server. If the value of a header field is an empty string then this field will not be passed to a proxied server.

It’s also important to distinguish between request headers and response headers. Request headers are for traffic inbound to the webserver or backend app. Response headers are going the other way (in the HTTP response you get back using client, e.g. curl or browser).

Ok, so look at the following short explanation about proxy directives (for more information about valid header values please see this rule):

• proxy_http_version — defines the HTTP protocol version for proxying, by default it it set to 1.0. For Websockets and keepalive connections you need to use the version 1.1:

• proxy_cache_bypass — sets conditions under which the response will not be taken from a cache:

  proxy_cache_bypass $http_upgrade;

• proxy_intercept_errors — means that any response with HTTP code 300 or greater is handled by the error_page directive and ensures that if the proxied backend returns an error status, NGINX will be the one showing the error page (as opposed to the error page on the backend side). If you want certain error pages still being delivered from the upstream server, then simply don’t specify the error_page <code> on the reverse proxy (without this, NGINX will forward the error page coming from the upstream server to the client):

  proxy_intercept_errors on;
  error_page 404 /404.html; # from proxy
  
  # To bypass error intercepting (if you have proxy_intercept_errors on):
  # 1 - don't specify the error_page 404 on the reverse proxy
  # 2 - go to the @debug location
  error_page 500 503 504 @debug;
  location @debug {
    proxy_intercept_errors off;
    proxy_pass http://backend;
  }

• proxy_set_header — allows redefining or appending fields to the request header passed to the proxied server

  • Upgrade and Connection — these header fields are required if your application is using Websockets:

    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";

  • Host — the $host variable in the following order of precedence contains: host name from the request line, or host name from the Host request header field, or the server name matching a request: NGINX uses Host header for server_name matching. It does not use TLS SNI. This means that for an SSL server, NGINX must be able to accept SSL connection, which boils down to having certificate/key. The cert/key can be any, e.g. self-signed:

    proxy_set_header Host $host;

  • X-Real-IP — forwards the real visitor remote IP address to the proxied server:

    proxy_set_header X-Real-IP $remote_addr;

  • X-Forwarded-For — is the conventional way of identifying the originating IP address of the user connecting to the web server coming from either a HTTP proxy or load balancer:

    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;

  • X-Forwarded-Proto — identifies the protocol (HTTP or HTTPS) that a client used to connect to your proxy or load balancer:

    proxy_set_header X-Forwarded-Proto $scheme;

  • X-Forwarded-Host — defines the original host requested by the client:

    proxy_set_header X-Forwarded-Host $host;

  • X-Forwarded-Port — defines the original port requested by the client:

    proxy_set_header X-Forwarded-Port $server_port;