This follows #8232 which added the code to generate the code required to parse
HCL files for each packer component.
All old config files of packer will keep on working the same. Packer takes one
argument. When a directory is passed, all files in the folder with a name
ending with “.pkr.hcl” or “.pkr.json” will be parsed using the HCL2 format.
When a file ending with “.pkr.hcl” or “.pkr.json” is passed it will be parsed
using the HCL2 format. For every other case; the old packer style will be used.
## 1. the hcl2template pkg can create a packer.Build from a set of HCL (v2) files
I had to make the packer.coreBuild (which is our one and only packer.Build ) a public struct with public fields
## 2. Components interfaces get a new ConfigSpec Method to read a file from an HCL file.
This is a breaking change for packer plugins.
a packer component can be a: builder/provisioner/post-processor
each component interface now gets a `ConfigSpec() hcldec.ObjectSpec`
which allows packer to tell what is the layout of the hcl2 config meant
to configure that specific component.
This ObjectSpec is sent through the wire (RPC) and a cty.Value is now
sent through the already existing configuration entrypoints:
Provisioner.Prepare(raws ...interface{}) error
Builder.Prepare(raws ...interface{}) ([]string, error)
PostProcessor.Configure(raws ...interface{}) error
close#1768
Example hcl files:
```hcl
// file amazon-ebs-kms-key/run.pkr.hcl
build {
sources = [
"source.amazon-ebs.first",
]
provisioner "shell" {
inline = [
"sleep 5"
]
}
post-processor "shell-local" {
inline = [
"sleep 5"
]
}
}
// amazon-ebs-kms-key/source.pkr.hcl
source "amazon-ebs" "first" {
ami_name = "hcl2-test"
region = "us-east-1"
instance_type = "t2.micro"
kms_key_id = "c729958f-c6ba-44cd-ab39-35ab68ce0a6c"
encrypt_boot = true
source_ami_filter {
filters {
virtualization-type = "hvm"
name = "amzn-ami-hvm-????.??.?.????????-x86_64-gp2"
root-device-type = "ebs"
}
most_recent = true
owners = ["amazon"]
}
launch_block_device_mappings {
device_name = "/dev/xvda"
volume_size = 20
volume_type = "gp2"
delete_on_termination = "true"
}
launch_block_device_mappings {
device_name = "/dev/xvdf"
volume_size = 500
volume_type = "gp2"
delete_on_termination = true
encrypted = true
}
ami_regions = ["eu-central-1"]
run_tags {
Name = "packer-solr-something"
stack-name = "DevOps Tools"
}
communicator = "ssh"
ssh_pty = true
ssh_username = "ec2-user"
associate_public_ip_address = true
}
```
It is simply the best/simplest solution and trying to prevent users from passing and integer here would be like opening a can of worms. Because:
* we cannot make mapstructure validate our duration string ( with an UnmarshalJSON func etc.)
* we cannot make mapstructure spit a string instead of a duration and packer will decode-encode-decode config.
* the hcl2 generated code asks for a string, so this will be enforced by default.
Before this commit it was possible to set a duration using an integer or a float. Go's time.Duration is an int64 internally an mapstructure will take advantage of this and load the number as a int64 but `1` means one ns which is unexpected/confusing. To avoid confusion and enforce readability this forces users to pass a string with a unit for a duration; ex "56s".
* make CustomerEncryptionKey our own type so that it can be hcl2 generated
* make Account setting unexported so that it doesn't temper with HCL2 generation ( the field is set a bit later after processing )
* removed packer.Cache and references since packer.Cache is never used except in the download step. The download step now uses the new func packer.CachePath(targetPath) for this, the behavior is the same.
* removed download code from packer that was reimplemented into the go-getter library: progress bar, http download restart, checksuming from file, skip already downloaded files, symlinking, make a download cancellable by context.
* on windows if packer is running without symlinking rights and we are getting a local file, the file will be copied instead to avoid errors.
* added unit tests for step_download that are now CI tested on windows, mac & linux.
* files are now downloaded under cache dir `sha1(filename + "?checksum=" + checksum) + file_extension`
* since the output dir is based on the source url and the checksum, when the checksum fails, the file is auto deleted.
* a download file is protected and locked by a file lock,
* updated docs
* updated go modules and vendors
If instance has a different DNS/search domain then "metadata" alone won't work. The FQDN of "metadata.google.internal" is included in hosts file, and will work regardless of DNS configuration. Alternatively the IP 169.254.169.254 can be used.
Because of deficiencies in the encrypted-PEM format, it's not always possible to detect an incorrect
password. In these cases no error will be returned but the decrypted DER bytes will be random noise.
this closes#3337
There were 5 different formats for the Packer useragent string. This
fixes that and unifies it into a helper package.
I did not touch oracle's user-agent, because it looked kinda special.
This change requires 'disable_default_service_account=false' in order to
set 'service_account_email'.
This is a guard against an incorrect assumption
that disabling the default service account would mean that no service
account would be used.
The ability to use a service account other than the default was
introduced in #5928. This change adds to that by introducing the
'disable_default_service_account' config option. If true - and
'service_account_email' is not set - Packer will create a GCE VM
with no service account.
This commit allows user to specify the service account they want
to associate with the virtual machine provisionned by setting
the service_account_email field in the config.
It allows to manage permissions of the instantiated VM properly,
using a service account that can be tied up to IAM roles and
permissions.
Run now takes a context as well as a statebag. We'll assign the context
to the blank identifier to prevent namespace collisions. We'll let the
step authors opt-in to using the context.
`find . -iname "step_*.go" -exec gsed -i'' 's/func \(.*\)Run(/func \1Run(_ context.Context, /' {} \;`
This change constructs partial URLs for networks and subnetworks if they
are not already partial or full URLs (i.e., they do not contain a '/' in
their name). Network and subnetwork self-links are no longer retrieved
from the API.
Previously, if a user did not provide the network or subnetwork as a
fully-qualified URL (i.e., self-link), the builder would make
compute.(sub)networks.get API calls with the provided identifier to
discover the self-link. This requires the user or service account Packer
is using to have permission to describe those network resources, which
is becoming less common as IAM is used more. Specifically, a user may
have permission to launch a VM into a network/subnetwork, but will not
have permission to call APIs to describe network resources.
If full server URL's is used in subnetwork we can skip reading from the network
API. This is usefull when you can launch instances in a shared network but don't
have access to do GET on the network resources.
Closes: #5018
Jeff Escalante
Megan Marsh
desolatorxxl
r_takaishi
Wilken Rivera
Rens Sikma
Sylvia Moss
Adrien Delorme
Adrien Delorme
Lars Lehtonen
Miles Crabill
Mason, Elliot
Pratyush singhal
krisko
Murcurio
Arnaud Dezandee
Mikhail Ushanov
Matthew Aynalem
deepuashokan85
DanHam
Richard Nienaber
Seth Vargo
Josh Soref
Evan Brown
Christophe Courtaut
Petr Hosek
Matthew Hooker
Daniel Hess
Atsushi Ishibashi
Peter Mounce
Patrick Decat
Dan Isla
Saulius Grusnys
Rickard von Essen
Dimitri Rudnev