This repository is an attempt to answer the age-old interview question "What happens when you type google.com into your browser's address box and press enter?"
Except instead of the usual story, we're going to try to answer this question in as much detail as possible. No skipping out on anything.
This is a collaborative process, so dig in and try to help out! There are tons of details missing, just waiting for you to add them! So send us a pull request, please!
This is all licensed under the terms of the Creative Commons Zero license.
Read this in 简体中文 (simplified Chinese), 日本語 (Japanese), 한국어 (Korean) and Spanish. NOTE: these have not been reviewed by the alex/what-happens-when maintainers.
- The "g" key is pressed
- The "Enter" key bottoms out
- Interrupt fires [NOT for USB keyboards]
- (On Windows) A
WM_KEYDOWN
message is sent to the app - (On OS X) A
KeyDown
NSEvent is sent to the app - (On GNU/Linux) the Xorg server listens for keycodes
- Parse URL
- Is it a URL or a search term?
- Convert non-ASCII Unicode characters in the hostname
- Check HSTS list
- DNS lookup
- ARP process
- Opening of a socket
- TLS handshake
- If a packet is dropped
- HTTP protocol
- HTTP Server Request Handle
- Behind the scenes of the Browser
- Browser
- HTML parsing
- CSS interpretation
- Page Rendering
- GPU Rendering
- Window Server
- Post-rendering and user-induced execution
The following sections explain the physical keyboard actions and the OS interrupts. When you press the key "g" the browser receives the event and the auto-complete functions kick in. Depending on your browser's algorithm and if you are in private/incognito mode or not various suggestions will be presented to you in the dropdown below the URL bar. Most of these algorithms sort and prioritize results based on search history, bookmarks, cookies, and popular searches from the internet as a whole. As you are typing "google.com" many blocks of code run and the suggestions will be refined with each keypress. It may even suggest "google.com" before you finish typing it.
To pick a zero point, let's choose the Enter key on the keyboard hitting the bottom of its range. At this point, an electrical circuit specific to the enter key is closed (either directly or capacitively). This allows a small amount of current to flow into the logic circuitry of the keyboard, which scans the state of each key switch, debounces the electrical noise of the rapid intermittent closure of the switch, and converts it to a keycode integer, in this case 13. The keyboard controller then encodes the keycode for transport to the computer. This is now almost universally over a Universal Serial Bus (USB) or Bluetooth connection, but historically has been over PS/2 or ADB connections.
In the case of the USB keyboard:
- The USB circuitry of the keyboard is powered by the 5V supply provided over pin 1 from the computer's USB host controller.
- The keycode generated is stored by internal keyboard circuitry memory in a register called "endpoint".
- The host USB controller polls that "endpoint" every ~10ms (minimum value declared by the keyboard), so it gets the keycode value stored on it.
- This value goes to the USB SIE (Serial Interface Engine) to be converted in one or more USB packets that follow the low-level USB protocol.
- Those packets are sent by a differential electrical signal over D+ and D- pins (the middle 2) at a maximum speed of 1.5 Mb/s, as an HID (Human Interface Device) device is always declared to be a "low-speed device" (USB 2.0 compliance).
- This serial signal is then decoded at the computer's host USB controller, and interpreted by the computer's Human Interface Device (HID) universal keyboard device driver. The value of the key is then passed into the operating system's hardware abstraction layer.
In the case of Virtual Keyboard (as in touch screen devices):
- When the user puts their finger on a modern capacitive touch screen, a
tiny amount of current gets transferred to the finger. This completes the
circuit through the electrostatic field of the conductive layer and
creates a voltage drop at that point on the screen. The
screen controller
then raises an interrupt reporting the coordinate of the keypress. - Then the mobile OS notifies the currently focused application of a press event in one of its GUI elements (which now is the virtual keyboard application buttons).
- The virtual keyboard can now raise a software interrupt for sending a 'key pressed' message back to the OS.
- This interrupt notifies the currently focused application of a 'key pressed' event.
The keyboard sends signals on its interrupt request line (IRQ), which is mapped
to an interrupt vector
(integer) by the interrupt controller. The CPU uses
the Interrupt Descriptor Table
(IDT) to map the interrupt vectors to
functions (interrupt handlers
) which are supplied by the kernel. When an
interrupt arrives, the CPU indexes the IDT with the interrupt vector and runs
the appropriate handler. Thus, the kernel is entered.
The HID transport passes the key down event to the KBDHID.sys
driver which
converts the HID usage into a scancode. In this case, the scan code is
VK_RETURN
(0x0D
). The KBDHID.sys
driver interfaces with the
KBDCLASS.sys
(keyboard class driver). This driver is responsible for
handling all keyboard and keypad input in a secure manner. It then calls into
Win32K.sys
(after potentially passing the message through 3rd party
keyboard filters that are installed). This all happens in kernel mode.
Win32K.sys
figures out what window is the active window through the
GetForegroundWindow()
API. This API provides the window handle of the
browser's address box. The main Windows "message pump" then calls
SendMessage(hWnd, WM_KEYDOWN, VK_RETURN, lParam)
. lParam
is a bitmask
that indicates further information about the keypress: repeat count (0 in this
case), the actual scan code (can be OEM dependent, but generally wouldn't be
for VK_RETURN
), whether extended keys (e.g. alt, shift, ctrl) were also
pressed (they weren't), and some other state.
The Windows SendMessage
API is a straightforward function that
adds the message to a queue for the particular window handle (hWnd
).
Later, the main message processing function (called a WindowProc
) assigned
to the hWnd
is called in order to process each message in the queue.
The window (hWnd
) that is active is actually an edit control and the
WindowProc
in this case has a message handler for WM_KEYDOWN
messages.
This code looks within the 3rd parameter that was passed to SendMessage
(wParam
) and, because it is VK_RETURN
knows the user has hit the ENTER
key.
The interrupt signal triggers an interrupt event in the I/O Kit kext keyboard
driver. The driver translates the signal into a key code which is passed to the
OS X WindowServer
process. Resultantly, the WindowServer
dispatches an
event to any appropriate (e.g. active or listening) applications through their
Mach port where it is placed into an event queue. Events can then be read from
this queue by threads with sufficient privileges calling the
mach_ipc_dispatch
function. This most commonly occurs through, and is
handled by, an NSApplication
main event loop, via an NSEvent
of
NSEventType
KeyDown
.
When a graphical X server
is used, X
will use the generic event
driver evdev
to acquire the keypress. A re-mapping of keycodes to scancodes
is made with X server
specific keymaps and rules.
When the scancode mapping of the key pressed is complete, the X server
sends the character to the window manager
(DWM, metacity, i3, etc), so the
window manager
in turn sends the character to the focused window.
The graphical API of the window that receives the character prints the
appropriate font symbol in the appropriate focused field.
The browser now has the following information contained in the URL (Uniform Resource Locator):
Protocol
"http"- Use 'Hyper Text Transfer Protocol'
Resource
"/"- Retrieve main (index) page
When no protocol or valid domain name is given the browser proceeds to feed the text given in the address box to the browser's default web search engine. In many cases the URL has a special piece of text appended to it to tell the search engine that it came from a particular browser's URL bar.
- The browser checks the hostname for characters that are not in
a-z
,A-Z
,0-9
,-
, or.
. - Since the hostname is
google.com
there won't be any, but if there were the browser would apply Punycode encoding to the hostname portion of the URL.
- The browser checks its "preloaded HSTS (HTTP Strict Transport Security)" list. This is a list of websites that have requested to be contacted via HTTPS only.
- If the website is in the list, the browser sends its request via HTTPS instead of HTTP. Otherwise, the initial request is sent via HTTP. (Note that a website can still use the HSTS policy without being in the HSTS list. The first HTTP request to the website by a user will receive a response requesting that the user only send HTTPS requests. However, this single HTTP request could potentially leave the user vulnerable to a downgrade attack, which is why the HSTS list is included in modern web browsers.)
- Browser checks if the domain is in its cache. (to see the DNS Cache in Chrome, go to chrome://net-internals/#dns).
- If not found, the browser calls
gethostbyname
library function (varies by OS) to do the lookup. gethostbyname
checks if the hostname can be resolved by reference in the localhosts
file (whose location varies by OS) before trying to resolve the hostname through DNS.- If
gethostbyname
does not have it cached nor can find it in thehosts
file then it makes a request to the DNS server configured in the network stack. This is typically the local router or the ISP's caching DNS server. - If the DNS server is on the same subnet the network library follows the
ARP process
below for the DNS server. - If the DNS server is on a different subnet, the network library follows
the
ARP process
below for the default gateway IP.
In order to send an ARP (Address Resolution Protocol) broadcast the network stack library needs the target IP address to lookup. It also needs to know the MAC address of the interface it will use to send out the ARP broadcast.
The ARP cache is first checked for an ARP entry for our target IP. If it is in the cache, the library function returns the result: Target IP = MAC.
If the entry is not in the ARP cache:
- The route table is looked up, to see if the Target IP address is on any of the subnets on the local route table. If it is, the library uses the interface associated with that subnet. If it is not, the library uses the interface that has the subnet of our default gateway.
- The MAC address of the selected network interface is looked up.
- The network library sends a Layer 2 (data link layer of the OSI model) ARP request:
ARP Request
:
Sender MAC: interface:mac:address:here Sender IP: interface.ip.goes.here Target MAC: FF:FF:FF:FF:FF:FF (Broadcast) Target IP: target.ip.goes.here
Depending on what type of hardware is between the computer and the router:
Directly connected:
- If the computer is directly connected to the router the router response
with an
ARP Reply
(see below)
Hub:
- If the computer is connected to a hub, the hub will broadcast the ARP
request out of all other ports. If the router is connected on the same "wire",
it will respond with an
ARP Reply
(see below).
Switch:
- If the computer is connected to a switch, the switch will check its local CAM/MAC table to see which port has the MAC address we are looking for. If the switch has no entry for the MAC address it will rebroadcast the ARP request to all other ports.
- If the switch has an entry in the MAC/CAM table it will send the ARP request to the port that has the MAC address we are looking for.
- If the router is on the same "wire", it will respond with an
ARP Reply
(see below)
ARP Reply
:
Sender MAC: target:mac:address:here Sender IP: target.ip.goes.here Target MAC: interface:mac:address:here Target IP: interface.ip.goes.here
Now that the network library has the IP address of either our DNS server or the default gateway it can resume its DNS process:
- The DNS client establishes a socket to UDP port 53 on the DNS server, using a source port above 1023.
- If the response size is too large, TCP will be used instead.
- If the local/ISP DNS server does not have it, then a recursive search is requested and that flows up the list of DNS servers until the SOA is reached, and if found an answer is returned.
Once the browser receives the IP address of the destination server, it takes
that and the given port number from the URL (the HTTP protocol defaults to port
80, and HTTPS to port 443), and makes a call to the system library function
named socket
and requests a TCP socket stream - AF_INET/AF_INET6
and
SOCK_STREAM
.
- This request is first passed to the Transport Layer where a TCP segment is crafted. The destination port is added to the header, and a source port is chosen from within the kernel's dynamic port range (ip_local_port_range in Linux).
- This segment is sent to the Network Layer, which wraps an additional IP header. The IP address of the destination server as well as that of the current machine is inserted to form a packet.
- The packet next arrives at the Link Layer. A frame header is added that includes the MAC address of the machine's NIC as well as the MAC address of the gateway (local router). As before, if the kernel does not know the MAC address of the gateway, it must broadcast an ARP query to find it.
At this point the packet is ready to be transmitted through either:
For most home or small business Internet connections the packet will pass from your computer, possibly through a local network, and then through a modem (MOdulator/DEModulator) which converts digital 1's and 0's into an analog signal suitable for transmission over telephone, cable, or wireless telephony connections. On the other end of the connection is another modem which converts the analog signal back into digital data to be processed by the next network node where the from and to addresses would be analyzed further.
Most larger businesses and some newer residential connections will have fiber or direct Ethernet connections in which case the data remains digital and is passed directly to the next network node for processing.
Eventually, the packet will reach the router managing the local subnet. From there, it will continue to travel to the autonomous system's (AS) border routers, other ASes, and finally to the destination server. Each router along the way extracts the destination address from the IP header and routes it to the appropriate next hop. The time to live (TTL) field in the IP header is decremented by one for each router that passes. The packet will be dropped if the TTL field reaches zero or if the current router has no space in its queue (perhaps due to network congestion).
This send and receive happens multiple times following the TCP connection flow:
- Client chooses an initial sequence number (ISN) and sends the packet to the server with the SYN bit set to indicate it is setting the ISN
- Server receives SYN and if it's in an agreeable mood:
- Server chooses its own initial sequence number
- Server sets SYN to indicate it is choosing its ISN
- Server copies the (client ISN +1) to its ACK field and adds the ACK flag to indicate it is acknowledging receipt of the first packet
- Client acknowledges the connection by sending a packet:
- Increases its own sequence number
- Increases the receiver acknowledgment number
- Sets ACK field
- Data is transferred as follows:
- As one side sends N data bytes, it increases its SEQ by that number
- When the other side acknowledges receipt of that packet (or a string of packets), it sends an ACK packet with the ACK value equal to the last received sequence from the other
- To close the connection:
- The closer sends a FIN packet
- The other sides ACKs the FIN packet and sends its own FIN
- The closer acknowledges the other side's FIN with an ACK
- The client computer sends a
ClientHello
message to the server with its Transport Layer Security (TLS) version, list of cipher algorithms and compression methods available. - The server replies with a
ServerHello
message to the client with the TLS version, selected cipher, selected compression methods and the server's public certificate signed by a CA (Certificate Authority). The certificate contains a public key that will be used by the client to encrypt the rest of the handshake until a symmetric key can be agreed upon. - The client verifies the server digital certificate against its list of trusted CAs. If trust can be established based on the CA, the client generates a string of pseudo-random bytes and encrypts this with the server's public key. These random bytes can be used to determine the symmetric key.
- The server decrypts the random bytes using its private key and uses these bytes to generate its own copy of the symmetric master key.
- The client sends a
Finished
message to the server, encrypting a hash of the transmission up to this point with the symmetric key. - The server generates its own hash, and then decrypts the client-sent hash
to verify that it matches. If it does, it sends its own
Finished
message to the client, also encrypted with the symmetric key. - From now on the TLS session transmits the application (HTTP) data encrypted with the agreed symmetric key.
Sometimes, due to network congestion or flaky hardware connections, TLS packets will be dropped before they get to their final destination. The sender then has to decide how to react. The algorithm for this is called TCP congestion control. This varies depending on the sender; the most common algorithms are cubic on newer operating systems and New Reno on almost all others.
- Client chooses a congestion window based on the maximum segment size (MSS) of the connection.
- For each packet acknowledged, the window doubles in size until it reaches the 'slow-start threshold'. In some implementations, this threshold is adaptive.
- After reaching the slow-start threshold, the window increases additively for each packet acknowledged. If a packet is dropped, the window reduces exponentially until another packet is acknowledged.
If the web browser used was written by Google, instead of sending an HTTP request to retrieve the page, it will send a request to try and negotiate with the server an "upgrade" from HTTP to the SPDY protocol.
If the client is using the HTTP protocol and does not support SPDY, it sends a request to the server of the form:
GET / HTTP/1.1 Host: google.com Connection: close [other headers]
where [other headers]
refers to a series of colon-separated key-value pairs
formatted as per the HTTP specification and separated by single newlines.
(This assumes the web browser being used doesn't have any bugs violating the
HTTP spec. This also assumes that the web browser is using HTTP/1.1
,
otherwise it may not include the Host
header in the request and the version
specified in the GET
request will either be HTTP/1.0
or HTTP/0.9
.)
HTTP/1.1 defines the "close" connection option for the sender to signal that the connection will be closed after completion of the response. For example,
Connection: close
HTTP/1.1 applications that do not support persistent connections MUST include the "close" connection option in every message.
After sending the request and headers, the web browser sends a single blank newline to the server indicating that the content of the request is done.
The server responds with a response code denoting the status of the request and responds with a response of the form:
200 OK [response headers]
Followed by a single newline, and then sends a payload of the HTML content of
www.google.com
. The server may then either close the connection, or if
headers sent by the client requested it, keep the connection open to be reused
for further requests.
If the HTTP headers sent by the web browser included sufficient information for
the webserver to determine if the version of the file cached by the web
browser has been unmodified since the last retrieval (ie. if the web browser
included an ETag
header), it may instead respond with a request of
the form:
304 Not Modified [response headers]
and no payload, and the web browser instead retrieve the HTML from its cache.
After parsing the HTML, the web browser (and server) repeats this process
for every resource (image, CSS, favicon.ico, etc) referenced by the HTML page,
except instead of GET / HTTP/1.1
the request will be
GET /$(URL relative to www.google.com) HTTP/1.1
.
If the HTML referenced a resource on a different domain than
www.google.com
, the web browser goes back to the steps involved in
resolving the other domain, and follows all steps up to this point for that
domain. The Host
header in the request will be set to the appropriate
server name instead of google.com
.
The HTTPD (HTTP Daemon) server is the one handling the requests/responses on the server-side. The most common HTTPD servers are Apache or nginx for Linux and IIS for Windows.
- The HTTPD (HTTP Daemon) receives the request.
- The server breaks down the request to the following parameters:
- HTTP Request Method (either
GET
,HEAD
,POST
,PUT
,PATCH
,DELETE
,CONNECT
,OPTIONS
, orTRACE
). In the case of a URL entered directly into the address bar, this will beGET
. - Domain, in this case - google.com.
- Requested path/page, in this case - / (as no specific path/page was requested, / is the default path).
- HTTP Request Method (either
- The server verifies that there is a Virtual Host configured on the server that corresponds with google.com.
- The server verifies that google.com can accept GET requests.
- The server verifies that the client is allowed to use this method (by IP, authentication, etc.).
- If the server has a rewrite module installed (like mod_rewrite for Apache or URL Rewrite for IIS), it tries to match the request against one of the configured rules. If a matching rule is found, the server uses that rule to rewrite the request.
- The server goes to pull the content that corresponds with the request, in our case it will fall back to the index file, as "/" is the main file (some cases can override this, but this is the most common method).
- The server parses the file according to the handler. If Google is running on PHP, the server uses PHP to interpret the index file, and streams the output to the client.
Once the server supplies the resources (HTML, CSS, JS, images, etc.) to the browser it undergoes the below process:
- Parsing - HTML, CSS, JS
- Rendering - Construct DOM Tree → Render Tree → Layout of Render Tree → Painting the render tree
The browser's functionality is to present the web resource you choose, by requesting it from the server and displaying it in the browser window. The resource is usually an HTML document, but may also be a PDF, image, or some other type of content. The location of the resource is specified by the user using a URI (Uniform Resource Identifier).
The way the browser interprets and displays HTML files is specified in the HTML and CSS specifications. These specifications are maintained by the W3C (World Wide Web Consortium) organization, which is the standards organization for the web.
Browser user interfaces have a lot in common with each other. Among the common user interface elements are:
- An address bar for inserting a URI
- Back and forward buttons
- Bookmarking options
- Refresh and stop buttons for refreshing or stopping the loading of current documents
- Home button that takes you to your home page
Browser High-Level Structure
The components of the browsers are:
- User interface: The user interface includes the address bar, back/forward button, bookmarking menu, etc. Every part of the browser display except the window where you see the requested page.
- Browser engine: The browser engine marshals actions between the UI and the rendering engine.
- Rendering engine: The rendering engine is responsible for displaying requested content. For example if the requested content is HTML, the rendering engine parses HTML and CSS, and displays the parsed content on the screen.
- Networking: The networking handles network calls such as HTTP requests, using different implementations for different platforms behind a platform-independent interface.
- UI backend: The UI backend is used for drawing basic widgets like combo boxes and windows. This backend exposes a generic interface that is not platform-specific. Underneath it uses operating system user interface methods.
- JavaScript engine: The JavaScript engine is used to parse and execute JavaScript code.
- Data storage: The data storage is a persistence layer. The browser may need to save all sorts of data locally, such as cookies. Browsers also support storage mechanisms such as localStorage, IndexedDB, WebSQL and FileSystem.
The rendering engine starts getting the contents of the requested document from the networking layer. This will usually be done in 8kB chunks.
The primary job of the HTML parser is to parse the HTML markup into a parse tree.
The output tree (the "parse tree") is a tree of DOM element and attribute nodes. DOM is short for Document Object Model. It is the object presentation of the HTML document and the interface of HTML elements to the outside world like JavaScript. The root of the tree is the "Document" object. Prior to any manipulation via scripting, the DOM has an almost one-to-one relation to the markup.
The parsing algorithm
HTML cannot be parsed using the regular top-down or bottom-up parsers.
The reasons are:
- The forgiving nature of the language.
- The fact that browsers have traditional error tolerance to support well known cases of invalid HTML.
- The parsing process is reentrant. For other languages, the source doesn't change during parsing, but in HTML, dynamic code (such as script elements containing document.write() calls) can add extra tokens, so the parsing process actually modifies the input.
Unable to use the regular parsing techniques, the browser utilizes a custom parser for parsing HTML. The parsing algorithm is described in detail by the HTML5 specification.
The algorithm consists of two stages: tokenization and tree construction.
Actions when the parsing is finished
The browser begins fetching external resources linked to the page (CSS, images, JavaScript files, etc.).
At this stage the browser marks the document as interactive and starts parsing scripts that are in "deferred" mode: those that should be executed after the document is parsed. The document state is set to "complete" and a "load" event is fired.
Note there is never an "Invalid Syntax" error on an HTML page. Browsers fix any invalid content and go on.
- Parse CSS files,
<style>
tag contents, andstyle
attribute values using "CSS lexical and syntax grammar" - Each CSS file is parsed into a
StyleSheet object
, where each object contains CSS rules with selectors and objects corresponding CSS grammar. - A CSS parser can be top-down or bottom-up when a specific parser generator is used.
- Create a 'Frame Tree' or 'Render Tree' by traversing the DOM nodes, and calculating the CSS style values for each node.
- Calculate the preferred width of each node in the 'Frame Tree' bottom-up by summing the preferred width of the child nodes and the node's horizontal margins, borders, and padding.
- Calculate the actual width of each node top-down by allocating each node's available width to its children.
- Calculate the height of each node bottom-up by applying text wrapping and summing the child node heights and the node's margins, borders, and padding.
- Calculate the coordinates of each node using the information calculated above.
- More complicated steps are taken when elements are
floated
, positionedabsolutely
orrelatively
, or other complex features are used. See http://dev.w3.org/csswg/css2/ and http://www.w3.org/Style/CSS/current-work for more details. - Create layers to describe which parts of the page can be animated as a group without being re-rasterized. Each frame/render object is assigned to a layer.
- Textures are allocated for each layer of the page.
- The frame/render objects for each layer are traversed and drawing commands are executed for their respective layer. This may be rasterized by the CPU or drawn on the GPU directly using D2D/SkiaGL.
- All of the above steps may reuse calculated values from the last time the webpage was rendered, so that incremental changes require less work.
- The page layers are sent to the compositing process where they are combined with layers for other visible content like the browser chrome, iframes and addon panels.
- Final layer positions are computed and the composite commands are issued via Direct3D/OpenGL. The GPU command buffer(s) are flushed to the GPU for asynchronous rendering and the frame is sent to the window server.
- During the rendering process the graphical computing layers can use general
purpose
CPU
or the graphical processorGPU
as well. - When using
GPU
for graphical rendering computations the graphical software layers split the task into multiple pieces, so it can take advantage ofGPU
massive parallelism for float point calculations required for the rendering process.
After rendering has been completed, the browser executes JavaScript code as a result of some timing mechanism (such as a Google Doodle animation) or user interaction (typing a query into the search box and receiving suggestions). Plugins such as Flash or Java may execute as well, although not at this time on the Google homepage. Scripts can cause additional network requests to be performed, as well as modify the page or its layout, causing another round of page rendering and painting.