Terraform 0.11 and prior had an odd special case where a resource
attribute access for "count" would be resolved as the count for the
whole resource, rather than as an attribute of an individual instance as
for all other attributes.
Because Terraform 0.12 makes test_instance.foo appear as a list when count
is set (so it can be used in other expressions), it's no longer possible
to have an attribute in that position: lists don't have attributes.
Fortunately we don't really need that special case anymore since it
doesn't do anything we can't now do with the length(...) function.
This upgrade rule, then, detects references like test_instance.foo.count
and rewrites to length(test_instance.foo). As a special case, if
test_instance.foo doesn't have "count" set then it just rewrites as the
constant 1, which mimics what would've happened in that case in Terraform
0.11.
Prior to Terraform v0.12 it was possible for a provider to secretly set
some default arguments for the "connection" block, which most commonly
included a hard-coded type of "ssh" and a value from "host".
In the interests of "explicit is better than implicit", Terraform 0.12 no
longer has this feature and instead requires connection settings to be
written explicitly in terms of the resource's exported attributes. For
compatibility though, the upgrade tool will insert expressions that are
as close as possible to the logic the provider formerly implemented, or
in a few rare cases a TF-UPGRADE-TODO comment to fix it up manually.
Some of the existing resource type implementations have incredibly
complicated implementations of selecting a single host IP address to use
and don't expose the result of that as an attribute, so for now we handle
those via a complicated Terraform language expression achieving the same
result. Ideally these providers would introduce a new attribute that
exports the same address formerly exported as the hostname before their
initial v0.12-compatible release, in which case we can simplify these to
just reference the attribute in question. That would be preferable also
because it would allow use of that exported attribute in other contexts,
such as in a null_resource provisioner somewhere else or in an output
to allow a caller to deal with the SSH part itself.
This uses the fixed "superset" schema from the main terraform package to
apply our standard expression mapping, with the exception of "type" where
interpolation sequences are not supported due to the type being evaluated
early to retrieve the schema for decoding the rest.
Aside from the two special meta-arguments "connection" and "provisioner"
this is just our standard mapping from schema to conversion rules, using
the provisioner's configuration schema.
Due to a copy-paste error, this was using the message from the providers
map in a "module" block.
This new message is not particularly helpful, but we should only see it
for a configuration that wouldn't have been valid in 0.11 either, and so
it's unlikely to be displayed.
Although sets do not have indexed elements, in Terraform 0.11 and earlier
element(...) would work with sets because we'd automatically convert them
to lists on entry to HIL -- with an arbitrary-but-consistent ordering --
and this return an arbitrary-but-consistent element from the list.
The element(...) function in Terraform 0.12 does not allow this because it
is not safe in general, but there was an existing pattern relying on this
in Terraform 0.11 configs which this upgrade rule is intended to preserve:
resource "example" "example" {
count = "${length(any_set_attribute)}"
foo = "${element(any_set_attribute, count.index}"
}
The above works because the exact indices assigned in the conversion are
irrelevant: we're just asking Terraform to create one resource for each
distinct element in the set.
This upgrade rule therefore inserts an explicit conversion to list if it
is able to successfully provide that the given expression will return a
set type:
foo = "${element(tolist(any_set_attribute), count.index}"
This makes the conversion explicit, allowing users to decide if it is
safe and rework the configuration if not. Since our static type analysis
functionality focuses mainly on resource type attributes, in practice this
rule will only apply when the given expression is a statically-checkable
resource reference. Since sets are an SDK-only concept in Terraform 0.11
and earlier anyway, in practice that works out just right: it's not
possible for sets to appear anywhere else in older versions anyway.
The comma-separated syntax is now reserved only for object constructor
expressions in attribute values, so the upgrade tool rewrites block
arguments to be newline-separated instead.
This was already working but we didn't have an explicit test for it until
now.
Prior to Terraform 0.12, ignore_changes was implemented in a
flatmap-oriented fashion and so users found that they could (and in fact,
were often forced to) use the internal .% and .# suffixes flatmap uses to
ignore changes to the number of elements in a list or map.
Terraform 0.12 no longer uses that representation, so we'll interpret
ignoring changes to the length as ignoring changes to the entire
collection. While this is not a totally-equivalent change, in practice
this pattern was most often used in conjunction with specific keys from a
map in order to _effectively_ ignore the entire map, even though Terraform
didn't really support that.
HIL implemented its type conversions by rewriting its AST to include calls
to some undocumented builtin functions. Unfortunately those functions were
still explicitly callable if you could figure out the name for them, and
so they may have been used in the wild.
In particular, __builtin_StringToFloat was used as part of a workaround
for a HIL design flaw where it would prefer to convert strings to integers
rather than floats when performing arithmetic operations. This issue was,
indeed, the main reason for unifying int ant float into a single number
type in HCL. Since we published that as a suggested workaround, the
upgrade tool ought to fix it up.
The other cases have never been documented as a workaround, so they are
less likely to appear in the wild, but we might as well fix them up anyway
since we already have the conversion functions required to get the same
result in the new language.
To be safe/conservative, most of these convert to _two_ function calls
rather than just one, which ensures that these new expressions retain the
behavior of implicitly converting to the source type before running the
conversion. The new conversion functions only specify target type, and so
cannot guarantee identical results if the argument type does not exactly
match what was previously given as the parameter type in HIL.
HEREDOC tokens are a little more fussy than normal string sequences
because we need to preserve the whitespace within them along with the
start and end markers while we upgrade any interpolated expressions inside.
We need to do some work locally here because the HCL heredoc processing
"does too much" and throws away information we need to do a faithful
upgrade.
We also need to contend with the fact that Terraform <=0.11 had an older
version of HCL that accidentally permitted a degenerate form of heredoc
where the marker was at the end of the final line, like this:
degenerate = <<EOT
this should never have workedEOT
When we migrate this, we'll introduce the additional newline that is now
required, which will unfortunately slightly change the result string to
include a newline when parsed by 0.12, and so we'll need to call this out
as a caveat in the upgrade guide.
Since these error messages get printed in Terraform's output and we
encourage users to share them as part of bug reports, we should avoid
including sensitive information in them to reduce the risk of accidental
exposure.
Previously, configupgrade would panic if it encountered a HEREDOC. For
the time being, we will simply print out the HEREDOC as-is.
Unfortunately, we discovered that terraform 0.11's version of HCL
allowed for HEREDOCs with the termination delimiter inline (instead of
on a newline, which is technically correct). Since 0.12configupgrade
needs to be bug-compatible with terraform 0.11, we must roll back to the
same version of HCL used in terraform 0.11.
Objects with DynamicPseudoType attributes can't be coerced within a map
if a concrete type is set. Change the Value type used to an Object when
there is a type mismatch.
In prior versions, we recommended using hash functions in conjunction with
the file function as an idiom for detecting changes to upstream blobs
without fetching and comparing the whole blob.
That approach relied on us being able to return raw binary data from
file(...). Since Terraform strings pass through intermediate
representations that are not binary-safe (e.g. the JSON state), there was
a risk of string corruption in prior versions which we have avoided for
0.12 by requiring that file(...) be used only with UTF-8 text files.
The specific case of returning a string and immediately passing it into
another function was not actually subject to that corruption risk, since
the HIL interpreter would just pass the string through verbatim, but this
is still now forbidden as a result of the stricter handling of file(...).
To avoid breaking these use-cases, here we introduce variants of the hash
functions a with "file" prefix that take a filename for a disk file to
hash rather than hashing the given string directly. The configuration
upgrade tool also now includes a rule to detect the documented idiom and
rewrite it into a single function call for one of these new functions.
This does cause a bit of function sprawl, but that seems preferable to
introducing more complex rules for when file(...) can and cannot read
binary files, making the behavior of these various functions easier to
understand in isolation.
There are a few constructs from 0.11 and prior that cause 0.12 parsing to
fail altogether, which previously created a chicken/egg problem because
we need to install the providers in order to run "terraform 0.12upgrade"
and thus fix the problem.
This changes "terraform init" to use the new "early configuration" loader
for module and provider installation. This is built on the more permissive
parser in the terraform-config-inspect package, and so it allows us to
read out the top-level blocks from the configuration while accepting
legacy HCL syntax.
In the long run this will let us do version compatibility detection before
attempting a "real" config load, giving us better error messages for any
future syntax additions, but in the short term the key thing is that it
allows us to install the dependencies even if the configuration isn't
fully valid.
Because backend init still requires full configuration, this introduces a
new mode of terraform init where it detects heuristically if it seems like
we need to do a configuration upgrade and does a partial init if so,
before finally directing the user to run "terraform 0.12upgrade" before
running any other commands.
The heuristic here is based on two assumptions:
- If the "early" loader finds no errors but the normal loader does, the
configuration is likely to be valid for Terraform 0.11 but not 0.12.
- If there's already a version constraint in the configuration that
excludes Terraform versions prior to v0.12 then the configuration is
probably _already_ upgraded and so it's just a normal syntax error,
even if the early loader didn't detect it.
Once the upgrade process is removed in 0.13.0 (users will be required to
go stepwise 0.11 -> 0.12 -> 0.13 to upgrade after that), some of this can
be simplified to remove that special mode, but the idea of doing the
dependency version checks against the liberal parser will remain valuable
to increase our chances of reporting version-based incompatibilities
rather than syntax errors as we add new features in future.
The parent commit fixes an issue where this would previously have led to
a crash. These new test cases verify that parsing is now able to complete
without crashing, though the result is still invalid.
In early versions of Terraform where the interpolation language didn't
have any real list support, list brackets around a single string was the
signal to split the string on a special uuid separator to produce a list
just in time for processing, giving expressions like this:
foo = ["${test_instance.foo.*.id}"]
Logically this is weird because it looks like it should produce a list
of lists of strings. When we added real list support in Terraform 0.7 we
retained support for this behavior by trimming off extra levels of list
during evaluation, and inadvertently continued relying on this notation
for correct type checking.
During the Terraform 0.10 line we fixed the type checker bugs (a few
remaining issues notwithstanding) so that it was finally possible to
use the more intuitive form:
foo = "${test_instance.foo.*.id}"
...but we continued trimming off extra levels of list for backward
compatibility.
Terraform 0.12 finally removes that compatibility shim, causing redundant
list brackets to be interpreted as a list of lists.
This upgrade rule attempts to identify situations that are relying on the
old compatibility behavior and trim off the redundant extra brackets. It's
not possible to do this fully-generally using only static analysis, but
we can gather enough information through or partial type inference
mechanism here to deal with the most common situations automatically and
produce a TF-UPGRADE-TODO comment for more complex scenarios where the
user intent isn't decidable with only static analysis.
In particular, this handles by far the most common situation of wrapping
list brackets around a splat expression like the first example above.
After this and the other upgrade rules are applied, the first example
above will become:
foo = test_instance.foo.*.id
By collecting information about the input variables during analysis, we
can return approximate type information for any references to those
variables in expressions.
Since Terraform 0.11 allowed maps of maps and lists of lists in certain
circumstances even though this was documented as forbidden, we
conservatively return collection types whose element types are unknown
here, which allows us to do shallow inference on them but will cause
us to get an incomplete result if any operations are performed on
elements of the list or map value.
Although we can't do fully-precise type inference with access only to a
single module's configuration, we can do some approximate inference using
some clues within the module along with our resource type schemas.
This depends on HCL's ability to pass through type information even if the
input values are unknown, mapping our partial input type information into
partial output type information by evaluating the same expressions.
This will allow us to do some upgrades that require dynamic analysis to
fully decide, by giving us three outcomes: needed, not needed, or unknown.
If it's unknown then that'll be our prompt to emit a warning for the user
to make a decision.
This is a temporary implementation of these rules just so that these can
be passed through verbatim (rather than generating an error) while we
do testing of other features.
A subsequent commit will finish these with their own custom rulesets.
The main tricky thing here is ignore_changes, which contains strings that
are better given as naked traversals in 0.12. We also handle here mapping
the old special case ["*"] value to the new "all" keyword.
Both resource blocks and module blocks contain references to providers
that are expressed as short-form provider addresses ("aws.foo" rather than
"provider.aws.foo").
These rules call for those to be unwrapped as naked identifiers during
upgrade, rather than appearing as quoted strings. This also introduces
some further rules for other simpler meta-arguments that are required
for the test fixtures for this feature.
Some further rules are required here to deal with the meta-arguments we
accept inside these blocks, but this is good enough to pass through most
module blocks using the standard attribute-expression-based mapping.
Previously we were handling this one as a special case, effectively
duplicating most of the logic from upgradeBlockBody.
By doing some prior analysis of the block we can produce a "rules" that
just passes through all of the attributes as-is, allowing us to reuse
upgradeBlockBody. This is a little weird for the locals block since
everything in it is user-selected names, but this facility will also be
useful in a future commit for dealing with module blocks, which contain
a mixture of user-chosen and reserved argument names.
We don't change JSON files at all and instead just emit a warning about
them since JSON files are usually generated rather than hand-written and
so any updates need to happen in the generator program rather than in its
output.
However, we do still need to copy them verbatim into the output map so
that we can keep track of them through any subsequent steps.
Prior to v0.12 Terraform was liberal about these and allowed them to
mismatch, but now it's important to get this right so that resources
and resource instances can be used directly as object values, and so
we'll fix up any sloppy existing references so things keep working as
expected.
This is particularly important for the pattern of using count to create
conditional resources, since previously the "true" case would create one
instance and Terraform would accept an unindexed reference to that.
The reference syntax is not significantly changed, but there are some
minor additional restrictions on identifiers in HCL2 and as a special case
we need to rewrite references to data.terraform_remote_state .
Along with those mandatory upgrades, we will also switch references to
using normal index syntax where it's safe to do so, as part of
de-emphasizing the old strange integer attribute syntax (like foo.0.bar).
Previously we were erroneously moving these out of their original block
into the surrounding body. Now we'll make sure to collect up any remaining
ad-hoc comments inside a nested block body before closing it.
Early on it looked like we wouldn't need to distinguish these since we
were only analyzing for provider types, but we're now leaning directly
on the resource addresses later on and so we need to make sure we produce
valid ones when data resources are present.
Users discovered that they could exploit some missing validation in
Terraform v0.11 and prior to treat block types as if they were attributes
and assign dynamic expressions to them, with some significant caveats and
gotchas resulting from the fact that this was never intended to work.
However, since such patterns are in use in the wild we'll convert them
to a dynamic block during upgrade. With only static analysis we must
unfortunately generate a very conservative, ugly dynamic block with
every possible argument set. Users ought to then clean up the generated
configuration after confirming which arguments are actually required.
We're using break elsewhere in here so it was weird to have a small set
of situations that return instead, which could then cause confusion for
future maintenance if a reader doesn't notice that control doesn't always
leave the outer switch statement.
If lookup is being used with only two arguments then it is equivalent to
index syntax and more readable that way, so we'll replace it.
Ideally we'd do similarly for element(...) here but sadly we cannot
because we can't prove in static analysis that the user is not relying
on the modulo wraparound behavior of that function.
We now have native language features for declaring tuples and objects,
which are the idiomatic way to construct sequence and mapping values that
can then be converted to list, set, and map types as needed.
In the old world, lists and maps could be created either using functions
in HIL or list/object constructs in HCL. Here we ensure that in the HCL
case we'll apply any required expression transformations to the individual
items within HCL's compound constructs.
Previously we were using the line count difference between the start of
one item and the next to decide whether to insert a blank line between
two items, but that is incorrect for multi-line items.
Instead, we'll now count the difference from the final line of the
previous item to the first line of the next, as best we can with the
limited position info recorded by the HCL1 parser.
The old parser was forgiving in allowing the use of block syntax where a
map attribute was expected, but the new parser is not (in order to allow
for dynamic map keys, for expressions, etc) and so the upgrade tool must
fix these to use attribute syntax.
The main area of interest in upgrading is dealing with special cases for
individual block items, so this generalization allows us to use the same
overall body-processing logic for everything but to specialize just how
individual items are dealt with, which we match by their names as given
in the original input source code.
This involved some refactoring of how block bodies are migrated, which
still needs some additional work to deal with meta-arguments but is now
at least partially generalized to support both resource and provider
blocks.
Martin Atkins