| Age | Commit message (Collapse) | Author |
|---|
|
This commit adds a check that all instructions in the relup are
valid. Earlier, the last part (after point_of_no_return) was not
checked before the actual installation started, meaning that if the
relup was hand written, there was a potential for crashing the upgrade
here.
|
|
Earlier, there was only one valid instruction to restart the emulator
in an appup/relup script, 'restart_new_emulator'. A new instuction,
'restart_emulator', is now added, and the meaning is as follows:
'restart_new_emulator' is mainly for the core applications (erts,
kernel, stdlib and sasl), and it indicates that there is new core
functionality in the release and the emulator needs to be restarted
before executing the upgrade scripts. If this instruction exists, a
temporary release will be created which consists of the new version
erts, kernel, stdlib and sasl, and the old versions of all other
applications. After restarting the emulator with this temporary
release, the rest of the upgrade instructions are executed, including
loading of new versions.
'restart_emulator' can be used by any application if a restart of the
emulator is needed after the upgrade instructions have been
executed. In this case, the emulator will be restarted with the new
release (i.e. not a tempoarary one) after all other upgrade
instructions for all applications have been excecuted.
|
|
If the version of erts differs between two releases, the release
handler automatically adds a 'restart_new_emulator' instruction to the
upgrade script (relup). Earlier, this instruction was always added at
the end of the upgrade script, causing the following sequence of
operations during an upgrade (a bit simplified):
1. suspend processes
2. load new code
3. execute code_change functions
4. resume processes
5. restart emulator with new erts version
Obviously, this caused the new code to be loaded into the old emulator
and this would fail if the beam format had been changed in the new
version of the emulator.
To overcome this problem, this commit changes the order of the
instructions, so for upgrade with changed erts version we now have:
1. restart emulator with new erts, kernel, stdlib and sasl versions,
but old versions of all other applications.
2. suspend processes
3. load new code
4. execute code_change functions
5. resume processes
This is implemented by creating a temporary release, including new
erts, kernel, stdlib and sasl from the new release and all other
applications from the old release. A new boot file for this temporary
release is created, which includes a new 'apply' instruction to run
release_handler:new_emulator_upgrade/2. Then the emulator is restarted
using this boot file - and release_handler:new_emulator_upgrade/2
executes the rest of the upgrade script.
For downgrade, the order will be as before:
1. suspend processes
2. execute code_change functions with 'down'-indication
3. load old code
4. resume processes
5. restart emulator with old erts version
|
|
* jw/release_handler_1:
General improvements to release_handler_1:get_supervised_procs
Conflicts:
lib/sasl/src/release_handler_1.erl
lib/sasl/test/release_handler_SUITE.erl
OTP-9546
|
|
The core issues this patch attempts to solve is two fold, 1) have
release_handler_1 act slightly differently in two corner cases and 2)
clean up the code in get_supervised_procs/0 to remove nested cases and
etc.
Regarding #1, get_supervised_procs/0 will now call functions to
test to see if the supervisor is suspended before attempting to ask it
for a list of children. It now will print an error message regarding
the suspended supervisor and produce an error that will cause the VM
to restart. Previously it would timeout attempting the call to
which_children and the VM would restart without any details regarding
the reason.
The second corner case is if in a child specification a supervisor is
incorrectly stated to be a worker and get_modules is called against it.
A timeout will occur causing a VM restart. Similar to the last corner
case in this patch an error message is printed and an error is emitted
causing a VM restart.
When first looking into the issue it was hard to discover why my
upgrades where failing. All I received during the upgrade process was
a timeout and a VM restart, no other information. This patch should
help users track down issues like these.
Regarding #2, due to the above confusion in trying to figure out what
had happened I dug into the code and started tracing it through and
found that the nested case statements and etc made it confusing. So I
started to rework and clean up, hopefully making this code path clearer
to future readers.
|
|
Given this option, all modules that are to be purged by indicated
upgrade,and that can be soft purged, will be purged when all other
check of check_install_release have been successfully completed.
I added a note under install_release in the reference manual about how
to use check_install_release with this new option in order to speed up
the execution of install_release.
I also added three more test cases for this functionality.
|
|
This commit utilizes the new bif erlang:check_old_code/1 to check if a
module has old code in the system or not before running
erlang:check_process_code/2. This is to optimize
release_handler:install_release (and
release_handler:check_install_release).
A new test is added which checks that after traversing all
processes/modules once and purging all old code, the second run
through this part of the code is "sufficiently" much faster.
A note is also added in the reference manual for
release_handler:install_release about how to go around the efficiency
problem of this function.
|
|
The appup instruction 'delete_module' would cause a crash during
upgrade if the module to be deleted was not loaded. The reason was
that the release_handler tried to read the version number of the old
module after the code path had changed to point to the new version of
the application. Thus, if the module had not been loaded before the
upgrade, there would no longer be any such module in the path
(delete_module indicates that the module is deleted from the new
version of the application).
This is corrected by letting the release_handler read the old version
of the module only if the module is updated - not if it is
removed. And it is always read before the code path is changed.
|
|
If a new version of an application did not include any erlang module
changes, the code path of the application was not updasted unless a
'load_object_code' instruction was added for the application. This
would be a problem if e.g. only some files in the priv dir were
changed, since calls to code:lib_dir or code:priv_dir would then point
to the old location of the application. This has been corrected - now
code:replace_path/2 will be called for all applications that are
changed (i.e. when the application's vsn is changed in the .rel file).
|
|
list
There are a couple of places in release_handler and release_handler_1
that assumed it has a disk to read from, which in the case of an
erl_prim_loader Loader other than efile is not necessarily true
Add check_paths/2 to do the equivalent of check_path/1 for when there
is a Masters list
Change get_vsn to no longer get sent File paths but instead use the
Bin since beam_lib:version being sent a file path causes it to read
the local file system
Add get_current_vsn/1 as an equivalent to
beam_lib:version(code:which(Mod)), but using erl_prim_loader:get_file
instead of reading from local file system
|
|
* sv/sasl-fix-get_status:
Prepare patch release
fix sys:get_status backward compatibility for sasl release_handler_1
OTP-8619 sv/sasl-fix-get_status
In R13B04 sys:get_status was modified to invoke format_status/2 in the
callback module if the module exports that function. This resulted in a
change to the term returned from calling sys:get_status on the supervisor
module, since supervisor is a gen_server and gen_server exports
format_status. The sasl release_handler_1 module had a dependency on the
pre-R13B04 term returned by sys:get_status when invoked on a supervisor, so
the R13B04 change broke that dependency.
This problem has been fixed by change release_handler_1 to handle both the
pre-R13B04 and R13B04 terms that sys:get_status can return from a
supervisor.
|
|
In R13B04 sys:get_status was modified to invoke format_status/2 in the
callback module if the module exports that function. This resulted in
a change to the term returned from calling sys:get_status on the
supervisor module, since supervisor is a gen_server and gen_server
exports format_status. The sasl release_handler_1 module had a
dependency on the pre-R13B04 term returned by sys:get_status when
invoked on a supervisor, so the R13B04 change broke that dependency.
Change release_handler_1 to handle both the pre-R13B04 and R13B04
terms that sys:get_status can return from a supervisor.
|
|
|
