blob: 39e085e2c803b6aa7dfb8107d8e6f66a61ac2156 [file] [log] [blame]
Name: sqlite
URL: http://sqlite.org/
Version: 3.8.7.4
Included In Release: Yes
Security Critical: Yes
License: Public domain
1) Managing differences between SQLite core and Chromium's version.
2) Making changes to Chromium SQLite.
3) Import new release of SQLite.
4) Running SQLite's test suite within the Chromium checkout.
---
1) Managing differences between SQLite core and Chromium's version.
Chromium maintains some differences WRT SQLite, for reasons beyond this
document's remit. Some differences are bugs we have found and fixed (and
hopefully upstreamed), some are fixes we've backported from a later version of
SQLite, and some our local changes unlikely to ever be upstreamed. New versions
of SQLite are imported every year or two, at which point the changes need to be
reviewed for continued applicability, and sometimes adjusted to reflect upstream
code changes.
To this end, the repository contains a reference copy of the SQLite source code
as of the last import, plus a series of patches which can be applied to
re-create the current trunk code. These patches are generated and processed by
git, with the intention of re-creating a changelist series so that importers can
use their regular revision-control knowledge to manage import merges.
---
2) Making changes to Chromium SQLite.
third_party/sqlite/src is the patched source from SQLite. This is used to
generate the amalgamation, a concatenation of all of the files into a giant
sqlite3.c. To prototype, edit in src/, then call
./google_generate_amalgamation.sh
to regenerate sqlite3.c. The code in src/ is much easier to edit, and the
SQLite test framework can easily be run. During development it may be
convenient to modify sqlite.gyp (or BUILD.gn) based on src/main.mk to just pull
in the src/ files rather than sqlite3.c.
Once your patch is complete, squash it down into a reasonable CL, then
re-generate the patches. This is a truncated version of the import flow. The
following is written like a shell script to allow copy/paste to a shell, ignore
comments and change the obvious lines. These instructions should work on Linux
or OSX. They may assume a modern version of git (I'm using 2.2.1).
# Everything based in sqlite subdir.
cd third_party/sqlite
BASE=3080704
#### Create a reference branch.
git checkout -b sqlite_${BASE} master
git rm -rf src
cp -a sqlite-src-${BASE} src
# -f includes ignored files, of which there are a couple.
git add -f src/
git commit -m "Reset to sqlite-src-${BASE}"
# This branch is unlikely to build.
#### Create a reference branch with patches applied.
git checkout -b sqlite_${BASE}_patched master
git rebase sqlite_${BASE}
git am --keep-non-patch patches/*.patch
git diff master
# This branch should be identical to master, unless someone forgot to export
# their changes into a patch. If so, do that as a separate CL and start over.
#### Cherry-pick your change.
git cherry-pick <your change>
# This branch should be identical to your development branch, except
# amalgamation.
# Rebuild the patch set.
git rm patches/*
git format-patch --output-directory=patches sqlite_${BASE}..HEAD
git add patches/*.patch
git commit -m "Rebuild patches for sqlite_${VERSION}"
# Re-generate the amalgamation.
./google_generate_amalgamation.sh
git commit -m 'google_generate_amalgamation.sh' amalgamation/
# At this point everything should build and work.
# Do a squash upload. This should add your single patch to patches/, and apply
# the changes your patch represents to src/ and amalgamation/. Other patches
# will have hash changes. A sensible check-in comment would be something like
# the patch's checkin comment, plus "regenerate amalgamation and generate patch
# file."
# TODO(shess): Should hash changes be checked in, or backed out?
# Find a sucker. Send review.
---
3) Import new release of SQLite.
Importing a new SQLite involves merging our local changes with SQLite's changes.
Like any other merge, this may involve dropping some CLs while modifying others.
The basic idea below is to generate git branches to work with:
sqlite_${BASE} - current version without patches
sqlite_${BASE}_patched - current version with patches applied via git CLs
sqlite_${VERSION} - new version without patches
sqlite_${VERSION}_patched - new version with patches applied via git CLs
At this point, squashing sqlite_${VERSION}_patched to master gives exactly a CL
suitable for committing.
# Everything based in sqlite subdir.
cd third_party/sqlite
BASE=3070603
VERSION=3080704
#### Create current-SQLite reference branch.
git checkout -b sqlite_${BASE} master
rm -rf src
cp -a sqlite-src-${BASE} src
# -f includes ignored files, of which there are a couple.
git add -f src/
git commit -m "Reset to sqlite-src-${BASE}"
# This branch is unlikely to build.
#### Convert patches into CLs.
git checkout -b sqlite_${BASE}_patched master
git rebase sqlite_${BASE}
git am --keep-non-patch patches/*.patch
git diff master
# This branch should be identical to master.
#### Create new-SQLite reference branch.
git checkout -b sqlite_${VERSION} master
git rebase sqlite_${BASE}
# SQLite's download page is at <http://www.sqlite.org/download.html>. Scroll to
# "Legacy Source Code Distribution Formats", and grab sqlite-src-<VERSION>.zip.
# Unzip it and pull it into the repo.
wget http://www.sqlite.org/2014/sqlite-src-${VERSION}.zip
unzip sqlite-src-${VERSION}.zip
rm sqlite-src-${VERSION}.zip
# -f includes ignored files, of which there are a couple.
git add -f sqlite-src-${VERSION}/
# Sometimes DOS line endings sneak into the source code. This command works on
# OSX and Linux and fixes those files, but double-check the results before
# committing:
egrep --exclude="*.eps" --exclude="*.ico" --exclude="*.jpg" \
--exclude="*.gif" --exclude="*.tiff" --exclude="*.vsix" -URl '\r' . | \
LANG=C xargs sed -i~ -e $'s/\r$//'
git commit -m "Begin import of sqlite-src-${VERSION}" sqlite-src-${VERSION}
rm -rf src
cp -a sqlite-src-${VERSION} src
# -f includes ignored files, of which there are a couple.
git add -f src/
git commit -m "Update src to sqlite-src-${VERSION}" src/
# This branch is unlikely to build.
#### Create a branch for merging the CLs to the new SQLite.
git checkout -b sqlite_${VERSION}_patched master
git rebase sqlite_${VERSION}
# Replay the patches onto this branch. There will be merge conflicts to fix.
# My approach is generally to just accept them prefering the patch's change in
# case of conflicts, and then resolve the conflicts as a second pass.
git rebase --onto zsqlite_${VERSION}_patched zsqlite_${BASE} zsqlite_${BASE}_patched
# Once everything is resolved, re-generate the amalgamation.
./google_generate_amalgamation.sh
git commit -a -m "google_generate_amalgamation.sh"
# The goal is to have a set of reasonably-independent CLs which can be
# understood separately, so that future importers can sensibly determine how to
# handle conflicts. So use git-rebase and slipstream fixups back into their
# original CL until things are relatively clean.
# Rebuild the patch set.
git rm patches/*
# This assumes that HEAD is still the google_generate_amalgamation.sh checkin.
git format-patch --output-directory=patches sqlite_${VERSION}..HEAD^
git add patches/*.patch
git commit -m "Rebuild patches for sqlite_${VERSION}"
# Drop the old version of SQLite.
git rm -r sqlite_${BASE}
git commit -m 'Remove sqlite_${BASE}' -- sqlite_${BASE}
# Do a squash upload. Edit the commit message appropriately to reflect anything
# from <http://www.sqlite.org/changes.html> which might be deemed important.
# Don't enumerate all of the patch messages, those are assumed, but do reference
# any material changes made.
# TODO(shess) Describe an appropriate comment style. Seems like it should at
# least include the SQLite version number. Meanwhile, look in the logs for past
# commits to model things on.
Find a sucker. Send review.
TODO(shess): It is basically impossible to trybot the entire change, it takes
forever to upload and sqlite3.c breaks things because it's too large. I have a
nasty Perl script to break up sqlite3.c into pieces which are then included by a
single .c file, but there must be a better way. Perhaps just have sqlite.gyp
include all the .c files directly?
Note that things can be broken down differently, if you prefer. For instance,
adding the new version of the SQLite distro and removing the old one can be
distinct CLs.
--------------------------------------------
4) Running SQLite's test suite within the Chromium checkout.
Prerequisites: The test suite requires tcl-dev and libicu-dev. Install those on
Ubuntu like:
sudo apt-get install tcl8.5-dev libicu-dev
On OSX, I use Homebrew:
sudo port install tcl
TODO(shess): OSX works fine with either tcl8.5 or tcl8.6, but on Ubuntu 14.04.1
with tcl8.6, I see crashes in some of the WAL tests. Revisit using tcl8.6 on
next import of SQLite.
TODO(shess): That Homebrew command has installed tcl8.6 for a few years, so the
above may require some adaptation of the build files.
cd third_party/sqlite/src
mkdir build
cd build
make -j -f ../Makefile.linux-gcc testfixture sqlite3
make -f ../Makefile.linux-gcc test > /tmp/test.log
egrep 'errors out of' /tmp/test.log
# Show broken tests:
egrep 'Failures on these tests:' /tmp/test.log
# Broken tests will also show lines ending in "..." instead of "... Ok".
In version 3.8.7.4 on OSX 10.9.5, I see:
6 errors out of 138390 tests
The failed tests are:
pager4-1.3 pager4-1.4 pager4-1.5 pager4-1.9 pager4-1.10 pager4-1.11
This is due to the change in os_unix.c fileHasMoved() to support WebDatabase.
Commenting out the early return allows them to succeed.
In version 3.8.7.4 on Ubuntu 14.04 I see:
9 errors out of 138920 tests
The failed tests are:
oserror-1.1.1 oserror-1.1.2 oserror-1.1.3
pager4-1.3 pager4-1.4 pager4-1.5 pager4-1.9 pager4-1.10 pager4-1.11
The oserror tests fail because there are too many fds available, and can be
fixed by running "ulimit -n 1024" before the test. The pager4 tests are failing
for the same reason as above.
--
NOTE(shess): On Ubuntu it is possible to run the tests in a tmpfs something
like:
TMPFS=/dev/shm/sqlite_build
BUILD=$PWD
mkdir $TMPFS
(cd $TMPFS ; $BUILD/testfixture $BUILD/../test/veryquick.test >/tmp/test.log)
This is faster, but it is plausible that different things are being tested than
real-world use.