This partially implements SQLite's bind-parameter expression to support
indicating placeholder values in a SQL statement. For example:
INSERT INTO table VALUES (42, ?);
In the above statement, the '?' identifier is a placeholder. This will
allow clients to compile statements a single time while running those
statements any number of times with different placeholder values.
Further, this will help mitigate SQL injection attacks.
This will make it easier to support both string types at the same time
while we convert code, and tracking down remaining uses.
One big exception is Value::to_string() in LibJS, where the name is
dictated by the ToString AO.
We have a new, improved string type coming up in AK (OOM aware, no null
state), and while it's going to use UTF-8, the name UTF8String is a
mouthful - so let's free up the String name by renaming the existing
class.
Making the old one have an annoying name will hopefully also help with
quick adoption :^)
Database::get_table currently either returns a RefPtr to an existing
table, a nullptr if the table doesn't exist, or an Error if some
internal error occured. Change this to return a NonnullRefPtr to an
exisiting table, or a SQL::Result with any error, including if the
table was not found. Callers can then handle that specific error code
if they want.
Returning a NonnullRefPtr will enable some further cleanup. This had
some fallout of needing to change some other methods' return types from
AK::ErrorOr to SQL::Result so that TRY may continue to be used.
Database::get_schema currently either returns a RefPtr to an existing
schema, a nullptr if the schema doesn't exist, or an Error if some
internal error occured. Change this to return a NonnullRefPtr to an
exisiting schema, or a SQL::Result with any error, including if the
schema was not found. Callers can then handle that specific error code
if they want.
Returning a NonnullRefPtr will enable some further cleanup. This had
some fallout of needing to change some other methods' return types from
AK::ErrorOr to SQL::Result so that TRY may continue to be used.
After splitting a node, the new node was written to the same pointer as
the current node - probably a copy / paste error. This new code requires
a `.pointer() -> u32` to exist on the object to be serialized,
preventing this issue from happening again.
Fixes#15844.
Currently, the Value class is essentially a "pImpl" wrapper around the
ValueImpl hierarchy of classes. This is a bit difficult to follow and
reason about, as methods jump between the Value class and its impl
classes.
This changes the Variant held by Value to instead store the specified
types (String, int, etc.) directly. In doing so, the ValueImpl classes
are removed, and all methods are now just concise Variant visitors.
As part of this rewrite, support for the "array" type is dropped (or
rather, just not re-implemented) as it was unused. If it's needed in the
future, support can be re-added.
This does retain the ability for non-NULL types to store NULL values
(i.e. an empty Optional). I tried dropping this support as well, but it
is depended upon by the on-disk storage classes in non-trivial ways.
Each of these strings would previously rely on StringView's char const*
constructor overload, which would call __builtin_strlen on the string.
Since we now have operator ""sv, we can replace these with much simpler
versions. This opens the door to being able to remove
StringView(char const*).
No functional changes.
Now that expression evaluation can use TRY, we can allow binary operator
methods to fail as well. This also fixes a few instances of converting a
Value to a double when we meant to convert to an integer.
The result of a SQL statement execution is either:
1. An error.
2. The list of rows inserted, deleted, selected, etc.
(2) is currently represented by a combination of the Result class and
the ResultSet list it holds. This worked okay, but issues start to
arise when trying to use Result in non-statement contexts (for example,
when introducing Result to SQL expression execution).
What we really need is for Result to be a thin wrapper that represents
both (1) and (2), and to not have any explicit members like a ResultSet.
So this commit removes ResultSet from Result, and introduces ResultOr,
which is just an alias for AK::ErrorOrr. Statement execution now returns
ResultOr<ResultSet> instead of Result. This further opens the door for
expression execution to return ResultOr<Value> in the future.
Lastly, this moves some other context held by Result over to ResultSet.
This includes the row count (which is really just the size of ResultSet)
and the command for which the result is for.
The implementation of LIKE uses regexes under the hood, and this
implementation of REGEXP takes the same approach. It employs
PosixExtended from LibRegex with case insensitive and Unicode flags
set. The implementation of LIKE is based on SQLlite specs, but SQLlite
does not offer directions for a built-in regex functionality, so this
one uses LibRegex.
Ordering is done by replacing the straight Vector holding the query
result in the SQLResult object with a dedicated Vector subclass that
inserts result rows according to their sort key using a binary search.
This is done in the ResultSet class.
There are limitations:
- "SELECT ... ORDER BY 1" (or 2 or 3 etc) is supposed to sort by the
n-th result column. This doesn't work yet
- "SELECT ... column-expression alias ... ORDER BY alias" is supposed to
sort by the column with the given alias. This doesn't work yet
What does work however is something like
```SELECT foo FROM bar SORT BY quux```
i.e. sorted by a column not in the result set. Once functions are
supported it should be possible to sort by random functions.
The evaluation order of method parameters is unspecified in C++, and
so we couldn't rely on parse_statement() being called before
parse_escape() when building a MatchExpression.
With this patch, we explicitly parse what we need in the right order,
before building the MatchExpression object.
Fixes a crash that was caused by a syntax error which is difficult to
catch by the parser: usually identifiers are accepted in column lists,
but they are not in a list of column values to be inserted in an INSERT.
Fixed this by putting in a heuristic check; we probably need a better
way to do this.
Included tests for this case.
Also introduced a new SQL Error code, `NotYetImplemented`, and return
that instead of crashing when encountering unimplemented SQL.
The handling of filesystem level errors was basically non-existing or
consisting of `VERIFY_NOT_REACHED` assertions. Addressed this by
* Adding `open` methods to `Heap` and `Database` which return errors.
* Changing the interface of methods of these classes and clients
downstream to propagate these errors.
The constructors of `Heap` and `Database` don't open the underlying
filesystem file anymore.
The SQL statement handlers return an `SQLErrorCode::InternalError`
error code if an error comes back from the lower levels. Note that some
of these errors are things like duplicate index entry errors that should
be caught before the SQL layer attempts to actually update the database.
Added tests to catch attempts to open weird or non-existent files as
databases.
Finally, in between me writing this patch and submitting the PR the
AK::Result<Foo, Bar> template got deprecated in favour of ErrorOr<Foo>.
This resulted in more busywork.
This patch introduces table joins. It uses a pretty dumb algorithm-
starting with a singleton '__unity__' row consisting of a single boolean
value, a cartesian product of all tables in the 'FROM' clause is built.
This cartesian product is then filtered through the 'WHERE' clause,
again without any smarts just using brute force.
This patch required a bunch of busy work to allow for example the
ColumnNameExpression having to deal with multiple tables potentially
having columns with the same name.
Because SQL is the craptastic language that it is, sometimes expressions
need to know details about the calling statement. For example the tables
in the 'FROM' clause may be needed to determine which columns are
referenced in 'WHERE' expressions. So the current statement is added
to the ExecutionContext and a new 'execute' overload on Statement is
created which takes the Database and the Statement and builds an
ExecutionContaxt from those.
These are needed to distinguish columns from different tables with the
same column name in one and the same (joined) Tuple. Not quite happy
yet with this API; I think some sort of hierarchical structure would be
better but we'll burn that bridge when we get there :^)
Command used:
grep -Pirn '(out|warn)ln\((?!["\)]|format,|stderr,|stdout,|output, ")' \
AK Kernel/ Tests/ Userland/
(Plus some manual reviewing.)
Let's pick ArgsParser as an example:
outln(file, m_general_help);
This will fail at runtime if the general help happens to contain braces.
Even if this transformation turns out to be unnecessary in a place or
two, this way the code is "more obviously" correct.
Our existing implementation did not check the element type of the other
pointer in the constructors and move assignment operators. This meant
that some operations that would require explicit casting on raw pointers
were done implicitly, such as:
- downcasting a base class to a derived class (e.g. `Kernel::Inode` =>
`Kernel::ProcFSDirectoryInode` in Kernel/ProcFS.cpp),
- casting to an unrelated type (e.g. `Promise<bool>` => `Promise<Empty>`
in LibIMAP/Client.cpp)
This, of course, allows gross violations of the type system, and makes
the need to type-check less obvious before downcasting. Luckily, while
adding the `static_ptr_cast`s, only two truly incorrect usages were
found; in the other instances, our casts just needed to be made
explicit.
Using a file(GLOB) to find all the test files in a directory is an easy
hack to get things started, but has some drawbacks. Namely, if you add
a test, it won't be found again without re-running CMake. `ninja` seems
to do this automatically, but it would be nice to one day stop seeing it
rechecking our globbed directories.
Classes reading and writing to the data heap would communicate directly
with the Heap object, and transfer ByteBuffers back and forth with it.
This makes things like caching and locking hard. Therefore all data
persistence activity will be funneled through a Serializer object which
in turn submits it to the Heap.
Introducing this unfortunately resulted in a huge amount of churn, in
which a number of smaller refactorings got caught up as well.
This patch provides very basic, bare bones implementations of the
INSERT and SELECT statements. They are *very* limited:
- The only variant of the INSERT statement that currently works is
SELECT INTO schema.table (column1, column2, ....) VALUES
(value11, value21, ...), (value12, value22, ...), ...
where the values are literals.
- The SELECT statement is even more limited, and is only provided to
allow verification of the INSERT statement. The only form implemented
is: SELECT * FROM schema.table
These statements required a bit of change in the Statement::execute
API. Originally execute only received a Database object as parameter.
This is not enough; we now pass an ExecutionContext object which
contains the Database, the current result set, and the last Tuple read
from the database. This object will undoubtedly evolve over time.
This API change dragged SQLServer::SQLStatement into the patch.
Another API addition is Expression::evaluate. This method is,
unsurprisingly, used to evaluate expressions, like the values in the
INSERT statement.
Finally, a new test file is added: TestSqlStatementExecution, which
tests the currently implemented statements. As the number and flavour of
implemented statements grows, this test file will probably have to be
restructured.
The implemtation of the Value class was based on lambda member variables
implementing type-dependent behaviour. This was done to ensure that
Values can be used as stack-only objects; the simplest alternative,
virtual methods, forces them onto the heap. The problem with the the
lambda approach is that it bloats the Values (which are supposed to be
lightweight objects) quite considerably, because every object contains
more than a dozen function pointers.
The solution to address both problems (we want Values to be able to live
on the stack and be as lightweight as possible) chosen here is to
encapsulate type-dependent behaviour and state in an implementation
class, and let the Value be an AK::Variant of those implementation
classes. All methods of Value are now basically straight delegates to
the implementation object using the Variant::visit method.
One issue complicating matters is the addition of two aggregate types,
Tuple and Array, which each contain a Vector of Values. At this point
Tuples and Arrays (and potential future aggregate types) can't contain
these aggregate types. This is limiting and needs to be addressed.
Another area that needs attention is the nomenclature of things; it's
a bit of a tangle of 'ValueBlahBlah' and 'ImplBlahBlah'. It makes sense
right now I think but admit we probably can do better.
Other things included here:
- Added the Boolean and Null types (and Tuple and Array, see above).
- to_string now always succeeds and returns a String instead of an
Optional. This had some impact on other sources.
- Added a lot of tests.
- Started moving the serialization mechanism more towards where I want
it to be, i.e. a 'DataSerializer' object which just takes
serialization and deserialization requests and knows for example how
to store long strings out-of-line.
One last remark: There is obviously a naming clash between the Tuple
class and the Tuple Value type. This is intentional; I plan to make the
Tuple class a subclass of Value (and hence Key and Row as well).
Since Clang enables a couple of warnings that we don't have in GCC,
these were not caught before. Included fixes:
- Use correct printf format string for `size_t`
- Don't compare Nonnull(Ref|Own)Ptr` to nullptr
- Fix unsigned int& => unsigned long& conversion
This patch introduces the SQLServer system server. This service is
supposed to be the only process/application talking to database storage.
This makes things like locking and caching more reliable, easier to
implement, and more efficient.
In LibSQL we added a client component that does the ugly IPC nitty-
gritty for you. All that's needed is setting a number of event handler
lambdas and you can connect to databases and execute statements on them.
Applications that wish to use this SQLClient class obviously need to
link LibSQL and LibIPC.
The Order enum is used in the Meta component of LibSQL. Using this enum
meant having to include the monster AST/AST.h include file. Furthermore,
they are sort of basic and therefore can live in the general SQL
namespace. Moved to LibSQL/Type.h.
Also introduced a new class, SQLResult, which is needed in future
patches.
Scanning tables is a linear process using pointers in the table's
tuples, and does not involve more 'stochastic' code paths like index
traversals. Therefore the 1000 and 10000 row tests were basically
overkill and added nothing we can't find out with less rows.
SQL was standardized before there was consensus on sane language syntax
constructs had evolved. The language is mostly case-insensitive, with
unquoted text converted to upper case. Identifiers can include lower
case characters and other 'special' characters by enclosing the
identifier with double quotes. A double quote is escaped by doubling it.
Likewise, a single quote in a literal string is escaped by doubling it.
All this means that the strategy used in the lexer, where a token's
value is a StringView 'window' on the source string, does not work,
because the value needs to be massaged before being handed to the
parser. Therefore a token now has a String containing its value. Given
the limited lifetime of a token, this is acceptable overhead.
Not doing this means that for example quote removal and double quote
escaping would need to be done in the parser or in AST node
construction, which would spread lexing basically all over the place.
Which would be suboptimal.
There was some impact on the sql utility and SyntaxHighlighter component
which was addressed by storing the token's end position together with
the start position in order to properly highlight it.
Finally, reviewing the tests for parsing numeric literals revealed an
inconsistency in which tokens we accept or reject: `1a` is accepted but
`1e` is rejected. Related to this is the fate of `0x`. Added a FIXME
reminding us to address this.
