Hi 


> > some minutes ago I published a PR for PostgreSQL support I developed
> > at INFN-Napoli for Darkside experiment...
> > 
> > I don't know if you receive a notification about this PR and in doubt
> > I wrote this message...
> 
> Hi, Gennaro, thank you for the very useful contribution. I saw the previous version 
> of your pull request and everything looked quite good. But that pull request was 
> for an older version of midas and it would not have applied cleanly to the current 
> version. I will take a look at your updated pull request. In theory it should only 
> add the Postgres class and modify a few other places in history_schema.cxx and have 
> no changes to anything else. (if you need those changes, it should be a separate 
> pull request).
> 
> Also I am curious what benefits and drawbacks of Postgres vs mysql/mariadb you have 
> observed for storing and using midas history data.
> 
> K.O.

Hi Konstantin,
thanks for this update |

My main interest for PostgreSQL is usage of TimescaleDB
(https://github.com/timescale/timescaledb) a PostgreSQL extension that
makes possible usage of downsampling functions on time-series...

here at INFN-Napoli we have a large history dataset that we manage
with MIDAS history and MySQL tables. We have a lot of issues 
(wait time, browser hangs, crashes) when we use MIDAS history plot 
pages on large time period because the Javascript web page try to 
download million of records in order to display them on a plot of 
(max) 2000 pixel width...

with native downsampling we can reduce a large dataset keeping the
"shape" of the curve using only the points needed by the plot area;

in TimescaleDB there is "lttb" ( Largest Triangle Three Bucket) a very 
nice and impressive downsampling function that preserve very well the
shape of the series.

If you are interested to see a lttb at work on some data you can open this page:
   https://www.base.is/flot

In next days I will work to add TimescaleDB backend to MIDAS history (it will be
similar to PostgreSQL backend) and we can discuss on how to add these 
downsampling features to history plot web pages, I already developed some
solutions and I will be happy to share them with MIDAS community;

Cheers,
Gennaro

> > some minutes ago I published a PR for PostgreSQL support I developed
> > at INFN-Napoli for Darkside experiment...
> > 
> > I don't know if you receive a notification about this PR and in doubt
> > I wrote this message...
> 
> Hi, Gennaro, thank you for the very useful contribution. I saw the previous version 
> of your pull request and everything looked quite good. But that pull request was 
> for an older version of midas and it would not have applied cleanly to the current 
> version. I will take a look at your updated pull request. In theory it should only 
> add the Postgres class and modify a few other places in history_schema.cxx and have 
> no changes to anything else. (if you need those changes, it should be a separate 
> pull request).
> 
> Also I am curious what benefits and drawbacks of Postgres vs mysql/mariadb you have 
> observed for storing and using midas history data.
> 
> K.O.

Hi,
I have updated the PR with a new one that includes TimescaleDB support and some
changes to mhistory.js to support downsampling queries...

Cheers,
Gennaro

> > some minutes ago I published a PR for PostgreSQL support I developed
> > at INFN-Napoli for Darkside experiment...
> > 
> > I don't know if you receive a notification about this PR and in doubt
> > I wrote this message...
> 
> Hi, Gennaro, thank you for the very useful contribution. I saw the previous version 
> of your pull request and everything looked quite good. But that pull request was 
> for an older version of midas and it would not have applied cleanly to the current 
> version. I will take a look at your updated pull request. In theory it should only 
> add the Postgres class and modify a few other places in history_schema.cxx and have 
> no changes to anything else. (if you need those changes, it should be a separate 
> pull request).
> 
> Also I am curious what benefits and drawbacks of Postgres vs mysql/mariadb you have 
> observed for storing and using midas history data.
> 
> K.O.

Hi,
is there any news regarding this pull request ? 
(https://bitbucket.org/tmidas/midas/pull-requests/30)

If you agree to merge I can resolve conflicts that now 
(after two months) are listed...

Regards,
Gennaro

> 
> Hi,
> I have updated the PR with a new one that includes TimescaleDB support and some
> changes to mhistory.js to support downsampling queries...
> 
> Cheers,
> Gennaro
> 
> > > some minutes ago I published a PR for PostgreSQL support I developed
> > > at INFN-Napoli for Darkside experiment...
> > > 
> > > I don't know if you receive a notification about this PR and in doubt
> > > I wrote this message...
> > 
> > Hi, Gennaro, thank you for the very useful contribution. I saw the previous version 
> > of your pull request and everything looked quite good. But that pull request was 
> > for an older version of midas and it would not have applied cleanly to the current 
> > version. I will take a look at your updated pull request. In theory it should only 
> > add the Postgres class and modify a few other places in history_schema.cxx and have 
> > no changes to anything else. (if you need those changes, it should be a separate 
> > pull request).
> > 
> > Also I am curious what benefits and drawbacks of Postgres vs mysql/mariadb you have 
> > observed for storing and using midas history data.
> > 
> > K.O.

> is there any news regarding this pull request ? 
> (https://bitbucket.org/tmidas/midas/pull-requests/30)

apologies for taking a very long time to review the proposed changes.

the main problem with this pull request remains, it tangles together too many changes to the code and I cannot simply 
say "this is okey", merge and commit it.

example of unrelated change is diff of mlogger.cxx, change of function in: "db_get_value(hDB, 0, "/Logger/Multithread 
transitions" ... )". there is also unrelated changes to whitespace sprinkled around.

can you review your diffs again and try to remove as much unrelated and unnecessary changes as you can?

I could do this for you, and merge my version, but next time you merge base midas, you will have a collision.

unrelated change of function is introduction of something called "downsampling", what is the purpose of this? How is it 
different from requesting binned data? Is it just a kludge to reduce the data size? Before we merge it, can you post a 
description/discussion to this forum here? (as a separate topic, separate from discussion of PostgreSQL merge).

the changes to add PostgreSQL so fat look reasonable:
- CMakeLists, is always painful but if you do same a MySQL, should be okey, we always end up rejigging this several 
times before it works everywhere.
- history.h, ok, minus changes for adding the "downsample" feature
- mlogger.cxx, changes are too tangled with "downsample" feature, cannot review
- SetDownsample() API is defective, should have separate Get() and Set() functions
- history_common.cxx, please do not add downsampling code to history providers that do not/will not support it.
- history_odbc.cxx, please do not change it. it does not support downsampling and never will.
- history.cxx, ditto
- mjsonrpc.cxx, history API is changed, we must know: is new JS compatible with old mhttpd? is old JS compatible with 
new mhttpd? (mixed versions are very common in practice). if there is incompatibility, can you recoded it to be 
compatible?
- history_schema.cxx: bitbucket diff is a dog's breakfast, cannot review. I will have to checkout your branch and diff 
by hand.

changes to mhistory.js appear to be extensive and some explanation is needed for what is changed, what bugs/problems 
are fixed, what new features are added.

to move forward, can you generate a pull requests that only adds pgsql to history_schema.cxx, history_common.cxx and 
mlogger.cxx and does not add any other functions, features and does not change any whistespace?

K.O.


> 
> If you agree to merge I can resolve conflicts that now 
> (after two months) are listed...
> 
> Regards,
> Gennaro
> 
> > 
> > Hi,
> > I have updated the PR with a new one that includes TimescaleDB support and some
> > changes to mhistory.js to support downsampling queries...
> > 
> > Cheers,
> > Gennaro
> > 
> > > > some minutes ago I published a PR for PostgreSQL support I developed
> > > > at INFN-Napoli for Darkside experiment...
> > > > 
> > > > I don't know if you receive a notification about this PR and in doubt
> > > > I wrote this message...
> > > 
> > > Hi, Gennaro, thank you for the very useful contribution. I saw the previous version 
> > > of your pull request and everything looked quite good. But that pull request was 
> > > for an older version of midas and it would not have applied cleanly to the current 
> > > version. I will take a look at your updated pull request. In theory it should only 
> > > add the Postgres class and modify a few other places in history_schema.cxx and have 
> > > no changes to anything else. (if you need those changes, it should be a separate 
> > > pull request).
> > > 
> > > Also I am curious what benefits and drawbacks of Postgres vs mysql/mariadb you have 
> > > observed for storing and using midas history data.
> > > 
> > > K.O.

> > is there any news regarding this pull request ? 
> > (https://bitbucket.org/tmidas/midas/pull-requests/30)
> 
> apologies for taking a very long time to review the proposed changes.
> 

I merged the PgSql bits by hand - the automatic tools make a dog's breakfast from the history_schema.cxx diffs. Ouch.

history_schema.cxx merged pretty much cleanly, but I have one question about CreateSqlColumn() with sql_strict set to "true". Can you say 
more why this is needed? Should this also be made the default for MySQL? The best I can tell the default values are only needed if we write 
to SQL but forget to provide values that should not be NULL? But our code never does this? Or this is for reading from SQL, where NULL values 
are replaced with the default values? I do not have time to look into this right now, I hope you can clarify it for me?

Also notice the fDownsample is set to zero and cannot be changed. I recommend we set it through the MakeMidasHistoryPgsql() factory method.

Please pull, merge, retest, update the pull request, check that there is no unrelated changes (changes in mlogger.cxx is a direct red flag!) 
and we should be able to merge the rest of your stuff pronto.

K.O.

commit e85bb6d37c85f02fc4895cae340ba71ab36de906 (HEAD -> develop, origin/develop, origin/HEAD)
Author: Konstantin Olchanski <olchansk@triumf.ca>
Date:   Fri Jul 21 09:45:08 2023 -0700

    merge PQSQL history in history_schema.cxx

commit f254ebd60a23c6ee2d4870f3b6b5e8e95a8f1f09
Author: Konstantin Olchanski <olchansk@triumf.ca>
Date:   Fri Jul 21 09:19:07 2023 -0700

    add PGSQL Makefile bits

commit aa5a35ba221c6f87ae7a811236881499e3d8dcf7
Author: Konstantin Olchanski <olchansk@triumf.ca>
Date:   Fri Jul 21 08:51:23 2023 -0700

    merge PGSQL support from https://bitbucket.org/gtortone/midas/branch/feature/timescaledb_support except for history_schema.cxx

Hi Konstantin,

thanks a lot for your work on PostgreSQL and TimescaleDB integration...
and sorry for unrelated changes on source code !

I will return on this task at end of this year (maybe October or November) because
I'm working on different tasks... but I will keep in mind your suggestions in order
to provide good source code.

Thanks,
Gennaro


> 
> I merged the PgSql bits by hand - the automatic tools make a dog's breakfast from the history_schema.cxx diffs. Ouch.
> 
> history_schema.cxx merged pretty much cleanly, but I have one question about CreateSqlColumn() with sql_strict set to "true". Can you say 
> more why this is needed? Should this also be made the default for MySQL? The best I can tell the default values are only needed if we write 
> to SQL but forget to provide values that should not be NULL? But our code never does this? Or this is for reading from SQL, where NULL values 
> are replaced with the default values? I do not have time to look into this right now, I hope you can clarify it for me?
> 
> Also notice the fDownsample is set to zero and cannot be changed. I recommend we set it through the MakeMidasHistoryPgsql() factory method.
> 
> Please pull, merge, retest, update the pull request, check that there is no unrelated changes (changes in mlogger.cxx is a direct red flag!) 
> and we should be able to merge the rest of your stuff pronto.
> 
> K.O.
> 
> commit e85bb6d37c85f02fc4895cae340ba71ab36de906 (HEAD -> develop, origin/develop, origin/HEAD)
> Author: Konstantin Olchanski <olchansk@triumf.ca>
> Date:   Fri Jul 21 09:45:08 2023 -0700
> 
>     merge PQSQL history in history_schema.cxx
> 
> commit f254ebd60a23c6ee2d4870f3b6b5e8e95a8f1f09
> Author: Konstantin Olchanski <olchansk@triumf.ca>
> Date:   Fri Jul 21 09:19:07 2023 -0700
> 
>     add PGSQL Makefile bits
> 
> commit aa5a35ba221c6f87ae7a811236881499e3d8dcf7
> Author: Konstantin Olchanski <olchansk@triumf.ca>
> Date:   Fri Jul 21 08:51:23 2023 -0700
> 
>     merge PGSQL support from https://bitbucket.org/gtortone/midas/branch/feature/timescaledb_support except for history_schema.cxx

The compilation of midas was broken by the last modification. The reason is that 

   Pgsql *fPgsql = NULL;

was not protected by 

#ifdef HAVE_PGSQL

So I put all PGSQL code under a big #ifdef and now it compiles again. You might want to double check my modification at

https://bitbucket.org/tmidas/midas/commits/e3c7e73459265e0d7d7a236669d1d1f2d9292a74

Best,
Stefan

> The compilation of midas was broken by the last modification. The reason is that 
>    Pgsql *fPgsql = NULL;
> was not protected by #ifdef HAVE_PGSQL

confirmed, my mistake, I forgot to test with "make cmake NO_PGSQL". your fix is correct, thanks.

K.O.

Hi!

I have released pymidas - Python binding to Midas.
It includes support for Online Database, Buffer, event
construction and parsing. 

We have used it for a couple years now here at CMD. (http://cmd.inp.nsk.su)
One of principal DAQ applications here (Slow Control Frontend) is
written in Python using pymidas.

http://cmd.inp.nsk.su/~bilenko/projects/pymidas/pymidas.html

Ciao, 
we have a very strange issue with python lib with client.odb_get("/") function 
when running as midas process and matplotlib is used.


we are developing a remote console by means of sending via kafka producer the odb, 
camera image and pmt waveforms, in the INFN cloud where grafana make available 
data for non expert shifters, as well as sending midas events for online 
reconstruction to the htcondr queue on cloud. The process work perfectly and allow 
use to parallelise to standard midas pipeline for file production, ecc the online 
monitoring and data processing where we have computing resources (our DAQ is 
underground at LNGS). Part of the work will be presented next weak at CHEP
the full code is available at https://github.com/CYGNUS-
RD/middleware/blob/master/dev/event_producer_s3.py
but to get the strange behaviour I report here a test script:

----
def main(verbose=False):
    from matplotlib import pyplot as plt

    import time

    import midas
    import midas.client

    
    client = midas.client.MidasClient("middleware")
    buffer_handle = client.open_event_buffer("SYSTEM",None,1000000000)
    request_id = client.register_event_request(buffer_handle, sampling_type = 2) 
    
    fpath = os.path.dirname(os.path.realpath(sys.argv[0]))
    
    
    while True:
        # 
        odb = client.odb_get("/")
        if verbose:
            print(odb)
        start1 = time.time()
              
        client.communicate(10)
        time.sleep(1)
        

    client.deregister_event_request(buffer_handle, request_id)

    client.disconnect()
----
if I run it as cli interactivity including or not matplotlib the everything si ok. 
As I run it as midas "program" I get: 
-----
Traceback (most recent call last):
  File "/home/standard/daq/middleware/dev/test_midas_error.py", line 48, in 
<module>
    main(verbose=options.verbose)
  File "/home/standard/daq/middleware/dev/test_midas_error.py", line 29, in main
    odb = client.odb_get("/")
  File "/home/standard/packages/midas/python/midas/client.py", line 354, in 
odb_get
    retval = midas.safe_to_json(buf.value, use_ordered_dict=True)
  File "/home/standard/packages/midas/python/midas/__init__.py", line 552, in 
safe_to_json
    return json.loads(decoded, strict=False, 
object_pairs_hook=collections.OrderedDict)
  File "/usr/lib/python3.8/json/__init__.py", line 370, in loads
    return cls(**kw).decode(s)
  File "/usr/lib/python3.8/json/decoder.py", line 337, in decode
    obj, end = self.raw_decode(s, idx=_w(s, 0).end())
  File "/usr/lib/python3.8/json/decoder.py", line 353, in raw_decode
    obj, end = self.scan_once(s, idx)
json.decoder.JSONDecodeError: Expecting property name enclosed in double quotes: 
line 300 column 26 (char 17535)
----
if I comment out the import of matplotlib every think works perfectly again also 
as midas program. 

it seams that there is a difference between the to way of use the code, and that 
is sufficient the call to matplotlib to corrupt in some way the odb. any ideas?

> it seams that there is a difference between the to way of use the code, and that 
> is sufficient the call to matplotlib to corrupt in some way the odb. any ideas?

I can't reproduce this on my machines, so this is going to be fun to debug!

Can you try running the program below please? It takes the important bits from odb_get() but prints out the string before we try to parse it as JSON. Feel free to send me the output via email (bsmith@triumf.ca) if you don't want to post your entire ODB dump in the elog.




import sys
import os
import time
import midas
import midas.client
import ctypes

def debug_get(client):
    c_path = ctypes.create_string_buffer(b"/")
    hKey = ctypes.c_int()
    client.lib.c_db_find_key(client.hDB, 0, c_path, ctypes.byref(hKey))

    buf = ctypes.c_char_p()
    bufsize = ctypes.c_int()
    bufend = ctypes.c_int()

    client.lib.c_db_copy_json_save(client.hDB, hKey, ctypes.byref(buf), ctypes.byref(bufsize), ctypes.byref(bufend))

    print("-" * 80)
    print("FULL DUMP")
    print("-" * 80)
    print(buf.value)
    print("-" * 80)
    print("Chars 17000-18000")
    print("-" * 80)
    print(buf.value[17000:18000])
    print("-" * 80)

    as_dict = midas.safe_to_json(buf.value, use_ordered_dict=True)

    client.lib.c_free(buf)

    return as_dict

def main(verbose=False):
    client = midas.client.MidasClient("middleware")
    buffer_handle = client.open_event_buffer("SYSTEM",None,1000000000)
    request_id = client.register_event_request(buffer_handle, sampling_type = 2)

    fpath = os.path.dirname(os.path.realpath(sys.argv[0]))

    while True:
        # odb = client.odb_get("/")
        odb = debug_get(client)

        if verbose:
            print(odb)
        start1 = time.time()

        client.communicate(10)
        time.sleep(1)


    client.deregister_event_request(buffer_handle, request_id)

    client.disconnect()

if __name__ == "__main__":
    main()

> > it seams that there is a difference between the to way of use the code, and that 
> > is sufficient the call to matplotlib to corrupt in some way the odb. any ideas?
> 
> I can't reproduce this on my machines, so this is going to be fun to debug!
> 
> Can you try running the program below please? It takes the important bits from odb_get() but prints out the string before we try to parse it as JSON. Feel free to send me the output via email (bsmith@triumf.ca) if you don't want to post your entire ODB dump in the elog.
> 
> 
> 
> 
> import sys
> import os
> import time
> import midas
> import midas.client
> import ctypes
> 
> def debug_get(client):
>     c_path = ctypes.create_string_buffer(b"/")
>     hKey = ctypes.c_int()
>     client.lib.c_db_find_key(client.hDB, 0, c_path, ctypes.byref(hKey))
> 
>     buf = ctypes.c_char_p()
>     bufsize = ctypes.c_int()
>     bufend = ctypes.c_int()
> 
>     client.lib.c_db_copy_json_save(client.hDB, hKey, ctypes.byref(buf), ctypes.byref(bufsize), ctypes.byref(bufend))
> 
>     print("-" * 80)
>     print("FULL DUMP")
>     print("-" * 80)
>     print(buf.value)
>     print("-" * 80)
>     print("Chars 17000-18000")
>     print("-" * 80)
>     print(buf.value[17000:18000])
>     print("-" * 80)
> 
>     as_dict = midas.safe_to_json(buf.value, use_ordered_dict=True)
> 
>     client.lib.c_free(buf)
> 
>     return as_dict
> 
> def main(verbose=False):
>     client = midas.client.MidasClient("middleware")
>     buffer_handle = client.open_event_buffer("SYSTEM",None,1000000000)
>     request_id = client.register_event_request(buffer_handle, sampling_type = 2)
> 
>     fpath = os.path.dirname(os.path.realpath(sys.argv[0]))
> 
>     while True:
>         # odb = client.odb_get("/")
>         odb = debug_get(client)
> 
>         if verbose:
>             print(odb)
>         start1 = time.time()
> 
>         client.communicate(10)
>         time.sleep(1)
> 
> 
>     client.deregister_event_request(buffer_handle, request_id)
> 
>     client.disconnect()
> 
> if __name__ == "__main__":
>     main()
Thank you!
if I added the mat

> > it seams that there is a difference between the to way of use the code, and that 
> > is sufficient the call to matplotlib to corrupt in some way the odb. any ideas?
> 
> I can't reproduce this on my machines, so this is going to be fun to debug!
> 
> Can you try running the program below please? It takes the important bits from odb_get() but prints out the string before we try to parse it as JSON. Feel free to send me the output via email (bsmith@triumf.ca) if you don't want to post your entire ODB dump in the elog.

Thank you!
if I added the matplotlib as follow:

#!/usr/bin/env python3

import sys
import os
import time
import midas
import midas.client
import ctypes
from matplotlib import pyplot as plt

def debug_get(client):
    c_path = ctypes.create_string_buffer(b"/")
    hKey = ctypes.c_int()
    client.lib.c_db_find_key(client.hDB, 0, c_path, ctypes.byref(hKey))

    buf = ctypes.c_char_p()
    bufsize = ctypes.c_int()
    bufend = ctypes.c_int()

    client.lib.c_db_copy_json_save(client.hDB, hKey, ctypes.byref(buf), ctypes.byref(bufsize), ctypes.byref(bufend))

    print("-" * 80)
    print("FULL DUMP")
    print("-" * 80)
    print(buf.value)
    print("-" * 80)
    print("Chars 17000-18000")
    print("-" * 80)
    print(buf.value[17000:18000])
    print("-" * 80)

    as_dict = midas.safe_to_json(buf.value, use_ordered_dict=True)

    client.lib.c_free(buf)

    return as_dict

def main(verbose=False):
    client = midas.client.MidasClient("middleware")
    buffer_handle = client.open_event_buffer("SYSTEM",None,1000000000)
    request_id = client.register_event_request(buffer_handle, sampling_type = 2)

    fpath = os.path.dirname(os.path.realpath(sys.argv[0]))

    while True:
        # odb = client.odb_get("/")
        odb = debug_get(client)

        if verbose:
            print(odb)
        start1 = time.time()

        client.communicate(10)
        time.sleep(1)


    client.deregister_event_request(buffer_handle, request_id)

    client.disconnect()

        
if __name__ == "__main__":
    from optparse import OptionParser
    parser = OptionParser(usage='usage: %prog\t ')
    parser.add_option('-v','--verbose', dest='verbose', action="store_true", default=False, help='verbose output;');
    (options, args) = parser.parse_args()
    main(verbose=options.verbose)


then tested the code in interactive mode without any error. as soon as I submit as midas "Program" I get the attached output.
thank you again, Giovanni

Looks like a localisation issue. Your floats are formatted as "6,6584e+01", whereas the JSON decoder expects "6.6584e+01".

Can you run the following few lines please? Then I'll be able to write a test using the same setup as you:


import locale
print(locale.getlocale())
from matplotlib import pyplot as plt
print(locale.getlocale())
 

> Looks like a localisation issue. Your floats are formatted as "6,6584e+01", whereas the JSON decoder expects "6.6584e+01".

This should be fixed in the latest commit to the midas develop branch. The JSON specification requires a dot for the decimal separator, so we must ignore the user's locale when formatting floats/doubles for JSON.

I've tested the fix on my machine by manually changing the locale, and also added an automated test in the python directory.

> > Looks like a localisation issue. Your floats are formatted as "6,6584e+01", whereas the JSON decoder expects "6.6584e+01".
> 
> This should be fixed in the latest commit to the midas develop branch. The JSON specification requires a dot for the decimal separator, so we must ignore the user's locale when formatting floats/doubles for JSON.
> 
> I've tested the fix on my machine by manually changing the locale, and also added an automated test in the python directory.

Thanks very macth Ben,
so if I understand correctly we have to update MIDAS to latest develop branch available? can you sand me the link to be sure of install the right update. 
can you also tell me how you fix manually? we are restarting and then well be difficult install and makes updete.
thank you again, regards, Giovanni

> 
> The only remaining problem when running my script is some kind of deadlock between the ODB and SYSMSG semaphores...
> 


In theory, we understand how programs that use 2 semaphores to protect 2 shared resources can deadlock
if there are mistakes in how locks are used.

For example, consider 2 semaphores A and B and 2 concurrent
subroutines foo() and bar() running at exactly the same time:

foo() { lock(A); lock(B); do stuff; unlock(B); unlock(A); } and
bar() { lock(B); lock(A); do stuff; unlock(A); unlock(B); }

This system will deadlock immediately with foo() taking semaphore A, bar() taking semaphore B,
then foo() waiting for B and bar() waiting for A forever.

This situation can also be described as a race condition where foo() and bar() are racing each
other to get the semaphores, with the result depending on who gets there first
and, in this case, sometimes the result is deadlock.

In this example, the size of the race condition time window is the wall clock time
between actually locking both semaphores in the sequence "lock(X); lock(Y);". While
locking a semaphore is "instantaneous", the actual function lock() takes time to call
and execute, and this time is not fixed - it can change if the CPU takes a hardware
interrupt (quick), a page fault (when we may have to wait until data is read from the swap file)
or a scheduler interrupt (when we are outright stopped for milliseconds while the CPU runs
some other process).

In reality, subroutines foo() and bar() do not run at exactly the same time, so the probability
of deadlock will depend on how often foo() and bar() are executed, the size of the race condition time window,
the number of processes executing foo() and bar(), and the amount of background activity
like swapping, hardware interrupts, etc.

(Also note that on a single-cpu system, we will probably never see a deadlock between foo() and bar()
because they will never be running at the same time. But the deadlock is still there, waiting
for the lucky moment when the scheduler switches from foo() to bar() just at the wrong place).

There is more on deadlocks and stuff written at:
http://en.wikipedia.org/wiki/Deadlock
http://en.wikipedia.org/wiki/Race_condition

In case of MIDAS, the 2 semaphores are the ODB lock and the SYSMSG lock (also remember about locks
for the shared memory event buffers, SYSTEM, etc, but they seem to be unlikely to deadlock).

The function foo() is any ODB function (db_xxx) that locks ODB and then calls cm_msg() (which locks SYSMSG).

The function bar() is cm_msg() which locks SYSMSG and then calls some ODB db_xxx() function which tries to lock ODB.

(This is made more interesting by cm_watchdog() periodically called by alarm(), where we alternately
take SYSMSG (via bm_cleanup) and ODB locks.)

I think this establishes a theoretical possibility for MIDAS to deadlock on the ODB and SYSMSG semaphores.

In practice, I think we almost never see this deadlock because cm_msg() is not called very often, and during normal
operation, is almost never called from inside ODB functions holding the ODB lock - almost all calls to cm_msg from
ODB functions are made to report some kind of problem with the ODB internal structure, something that "never"
happens.

By "luck" I stumbled into this deadlock when doing the "odbedit" fork-bomb torture tests, when high ODB lock
activity is combined with high cm_msg() activity reporting clients starting and stopping, combined with a large
number of MIDAS clients running, starting and stopping.

So a deadlock I see within 1 minute of running the torture test, other lucky people will see after running an experiment
for 1 year, or 1 month, or 1 day, depending.

In theory, this deadlock can be removed by establishing a fixed order of taking locks. There will never be a deadlock
if we always take the SYSMSG lock first, then ask for the ODB lock.

In practice, it means that using cm_msg() while holding an ODB lock is automatically dangerous
and should be avoided if not forbidden.

And it does work. By refactoring a few places in client startup, shutdown and cleanup code, I made the deadlock "go away",
and my test script (posted in my first message) no longer deadlocks, even if I run hundreds of odbedit's at the same time.

Unfortunately,  it is impractical to audit and refactor all of MIDAS to completely remove this problem. MIDAS call graphs
are sufficiently complicated for making manual analysis of lock sequences infeasible and
I expect any automatic lock analysis tool will be defeated by the cm_watchdog() periodic interrupt.

An improvement is possible if we make cm_msg() safe for calling from inside the ODB db_xxx() function. Instead
of immediately sending messages to SYSMSG (requiring a SYSMSG lock), if ODB is locked, cm_msg() could
save the messages in a buffer, which would be flushed when the ODB lock is released. (This does not fix
all the other places that take ODB and SYSMSG locks in arbitrary order, but I think those places are not as
likely to deadlock, compared to cm_msg()).

However, now that I have greatly reduced the probability of deadlock in the client startup/shutdown/cleanup code,
maybe there is no urgency for changing cm_msg() - remember that if we do not call cm_msg() we will never deadlock -
and during normal operation, cm_msg() is almost never called.

Investigation completed, I will now cleanup, retest and commit my changes to midas.c and odb.c. Looking into this
and writing it up was a good intellectual exercise.

P.S. Also remember that there are locks for shared memory event buffers (SYSTEM, etc), but those do not involve
lock inversion leading to deadlock. I think all lock sequences are like this: SYSTEM->ODB, SYSTEM->SYSMSG->ODB,
there are no inverted sequences SYSMSG->SYSTEM or ODB->SYSTEM and the only deadlocking
sequence SYSTEM->ODB->SYSMSG, does not really involve the SYSTEM lock.

K.O.

I really like the concept of the odbxx interface.
I think it would be a nice feature if one could have a read_only connection, e.g. by declaring a "const midas::odb".
Just for fun I tried if this already works, but the compiler doesn't allow const midas::odb for e.g. the [] operator. I'm guessing this would be non-trivial to implement, but I like the idea of certain Midas clients being able to read the odb without risking corruption.

> I really like the concept of the odbxx interface.
> I think it would be a nice feature if one could have a read_only connection, e.g. by declaring a "const midas::odb".
> Just for fun I tried if this already works, but the compiler doesn't allow const midas::odb for e.g. the [] operator. I'm guessing this would be non-trivial to implement, but I like the idea of certain Midas clients being able to read the odb without risking corruption.

Having a "const midas::odb" probably does not work (at least I would not know how to implement that).

But I could make an internal flag analog to the auto refresh flags. So you would have

    o.set_write_protect(true);

to turn on write protection. Would that work for you?

Best,
Stefan