> Fixed and committed. Can you check if it's working?
Yes, it is fixed. Thanks. K.O.

> I have noticed a problem with deleting a key to an array in odb, then
> recreating the record as in the code below. The record is recreated
> successfully, but when viewing it with mhttpd, a spurious blank line
> (coloured orange) is visible, followed by the rest of the data as normal.
> 
> db_create_record(hDB,0,"/Equipment/Cycle_scalers/Settings/",strcomb(type1_str));
>     }
>   else {
>     exp_mode = 2; /* TDmusr types - noscans */
>     status =
> db_create_record(hDB,0,"/Equipment/Cycle_scalers/Settings/",strcomb(type2_str));
>   }

The first problem is that the db_create_record has a trailing "/" in the key name 
after Settings. This causes the (empty) subsirectory which causes your trouble. 
Simple removing it fixes the problem. I agree that this is not obvious, so I 
added some code in db_create_record() which removes such a trailing slash if 
present. New version under CVS.

Second, the db_create_record() call is deprecated. You should use the new 
function db_check_record() instead, and remove your db_delete_key(). This avoids 
possible ODB trouble since the structure is not re-created each time, but only 
when necessary.

- Stefan

Fixed and committed. Can you check if it's working?

> fal.c does not compile: it calls cm_cleanup() with one argument when there
> should be two arguments. K.O.

Fixed and committed.

> > I have therefore implemented the function 
> > db_check_record(HNDLE hDB, HNDLE hKey, char *keyname, char *rec_str, BOOL
> correct)
> 
> Stephan, something is very wrong with the new code. My
> "/logger/channels/0/settings" is being destroyed on "begin run".

Okey. I found the problem in db_check_record(): when we decide that we have a
mismatch, we call db_create_record(...,rec_str), but by this time, rec_str no
longer points to the beginning of the ODB string because we started parsing it.

I tried this solution: save rec_str into rec_str_orig, then when we decide that
we have a mismatch, call db_create_record() with this saved rec_str_orig. It
fixes my immediate problem (destruction of "/logger/channels/0/settings"), but is
it correct?

I would like to fix it ASAP to get cvs-head working again: our mhttpd dumps core
on an assert() failure in db_create_record() and the set of db_check_record()
changes might fix it for me.

Here is the CVS diff:

RCS file: /usr/local/cvsroot/midas/src/odb.c,v
retrieving revision 1.73
diff -r1.73 odb.c
7810a7811
> char             *rec_str_orig = rec_str;
7820c7821
<     return db_create_record(hDB, hKey, keyname, rec_str);
---
>     return db_create_record(hDB, hKey, keyname, rec_str_orig);
7838c7839
<       return db_create_record(hDB, hKey, keyname, rec_str);
---
>       return db_create_record(hDB, hKey, keyname, rec_str_orig);
8023c8024
<               return db_create_record(hDB, hKey, keyname, rec_str);
---
>               return db_create_record(hDB, hKey, keyname, rec_str_orig);
8037c8038
<               return db_create_record(hDB, hKey, keyname, rec_str);
---
>               return db_create_record(hDB, hKey, keyname, rec_str_orig);

K.O.

fal.c does not compile: it calls cm_cleanup() with one argument when there
should be two arguments. K.O.

> I have therefore implemented the function 
> db_check_record(HNDLE hDB, HNDLE hKey, char *keyname, char *rec_str, BOOL
correct)

Stephan, something is very wrong with the new code. My
"/logger/channels/0/settings" is being destroyed on "begin run". Midas
checkout from october 31st is okey. This is a show stopper, but I am in a rush
and cannot debug it. I am falling back to the Oct 31st version... K.O.

Hi,

I have noticed a problem with deleting a key to an array in odb, then
recreating the record as in the code below. The record is recreated
successfully, but when viewing it with mhttpd, a spurious blank line
(coloured orange) is visible, followed by the rest of the data as normal.

This blank line causes trouble with experim.h because it
produces an empty structure e.g. :

#define CYCLE_SCALERS_SETTINGS_DEFINED

typedef struct {
  struct {
  } ;
  char      names[60][32];
} CYCLE_SCALERS_SETTINGS;

rather than :

#define CYCLE_SCALERS_SETTINGS_DEFINED

typedef struct {
  char      names[60][32];
} CYCLE_SCALERS_SETTINGS;


This empty structure causes a compilation error when rebuilding clients that
use experim.h

SD



 CYCLE_SCALERS_TYPE1_SETTINGS_STR(type1_str);
 CYCLE_SCALERS_TYPE2_SETTINGS_STR(type2_str);

Both type1_str and type2_str have been defined as in
experim.h
i.e.
#define CYCLE_SCALERS_TYPE1_SETTINGS_STR(_name) char *_name[] = {\
"[.]",\
"Names = STRING[60] :",\
"[32] Back%BSeg00",\
"[32] Back%BSeg01",\
 ........
 ........
"[32] General%NeutBm Cycle Sum",\
"[32] General%NeutBm Cycle Asym",\
"",\
NULL }

#define CYCLE_SCALERS_TYPE2_SETTINGS_STR(_name) char *_name[] = {\
"[.]",\
"Names = STRING[60] :",\
"[32] Back%BSeg00",\
"[32] Back%BSeg01",\
...........
............
"[32] General%B/F Cumul -",\
"[32] General%Asym Cumul -",\
"",\
NULL }

  if (db_find_key(hDB, 0, "/Equipment/Cycle_scalers/Settings/",&hKey) ==
DB_SUCCESS)
    db_delete_key(hDB,hKey,FALSE);
          
  if (  strncmp(fs.input.experiment_name,"1",1) == 0) {
      exp_mode = 1; /* Imusr type - scans */
      status =
db_create_record(hDB,0,"/Equipment/Cycle_scalers/Settings/",strcomb(type1_str));
    }
  else {
    exp_mode = 2; /* TDmusr types - noscans */
    status =
db_create_record(hDB,0,"/Equipment/Cycle_scalers/Settings/",strcomb(type2_str));
  }

> But there is one problem with "cleanup". It has a hardwired timeout of
> 2 seconds.  This is a problem for tasks like lazylogger which set a timeout
> of 60 seconds when moving the tape. So BEWARE, if you issue the "cleanup"
> command, it might kill some clients who have setup their timeout to longer
> than 2 seconds. 
> 
> I have asked Stefan to change this before. He said that, to be effective,
> the timeout value used for "cleanup" has to be rather short. 
> One possibility, would be to allow for a user entered "cleanup" timeout.
> The default could stay at 2 seconds. 

I changed the behaviour of cleanup by adding an extra parameter 
ignore_timeout to cm_cleanup(). Now, in ODBEdit, a "clanup" obeys the 
timeout set by the clients. The problem with that is if the logger crashes 
for example, and it's timeout is set o 5 min., it cannot be clean-up'ed any 
more for the next five minutes, and therefor not be restarted wasting 
precious beam time. That's why I hard-wired originally the "cleanup" timout 
to 2 sec. Now I added a flag "-f" to the ODBEdit cleanup command which works 
in the old fashion with a 2 sec. timeout. So a "cleanup" alone won't kill a 
looger which currently rewinds a tape or so, but a "cleanup -f" does.

I also changed internal timeouts from INT to DWORD, which should fix the 
problem Konstantin reported recently (re-starting an experiment after 
several weeks). New changes are commited, but I only did basic tests. So 
please try the new code and tell me if there is any problem.

- Stefan

Thanks for your contribution. Let me try to map your functionality to mvmestd calls:

> A) The VMIC vme_slave_xxx() options are not considered.

We could maybe do that through mvme_mmap(SLAVE, ...) instead of mvme_mmap(MASTER, ...)

> B) The interrupt handling can certainly match the 4 entries required in the user frontend
>     code i.e. Attach, Detach, Enable, Disable.

vmve_ioctl(VME_IOCTL_INTR_ATTACHE/DETACH/ENABLE/DISABLE, func())

> I don't understand your argument that the handle should be hidden. In case of multiple
> interfaces, how do you refer to a particular one if not specified? 
> The following scheme does require a handle for refering to the proper (device AND window).

Four reasons for that:

1) For the SBS/Bit3, you need a handle for each address mode. So if I have two crates (and I do in our 
current experiment), and have to access modules in A16, A24 and A32 mode, I need in total 6 handles. 
Sometimes I mix them up by mistake, and wonder why I get bus errors. 

2) Most installations will only have single crates (as your VMIC). So if there is only one crate, why 
bother with a handle? If you have hunderds of accesses in your code, you save some redundant typing work.

3) A handle is usually kept global, which is considered not good coding style.

4) Our MCSTD and MFBSTD functions also do not use a handle, so people used to those libraries will find it 
more natural not to use one.

> 1 ) deviceHandle = vme_init(int devNumber);
>     Even though the VMIC doesn't deal with multiple devices,
>     the SIS/PCI does and needs to init on a specific PCI card.
>     Internally:
>       opening of the device (/dev/sisxxxx_1) (ignored in case of VMIC).
>       Possible including a mapping to a default VME region of default size with default AM
>       (VMIC :16MB, A24). This way in a single call you get a valid handle for full VME access
>       in A24 mode. Needs to be elaborate this option. But in principle you need to declare the 
>      VME region that you want to work on (vme_map).

Just vme_init(); (like fb_init()).

This function takes the first device, opens it, and stores the handle internally. Sets the AM to a default 
value, and creates a mapping table which is initially empty or mapped to a default VME region. If one wants 
to access a secondary crate, one does a vme_ioctl(VME_IOCTL_CRATE_SET, 2), which opens the secondary crate, 
and stores the new handle in the internal table if applicable.

> 2) mapHandle = vme_map(int deviceHandle, int vmeAddress, int am, int size);
>     Return a mapHandle specific to a device and window. The am has to be specified.
>     What ever are the operation to get there, the mapHandle is a reference to thas setting.
>     It could just fill a map structure.
>     Internally:
>       WindowHandle[deviceHandle] = vme_master_create(BusHandle[deviceHandle], ...
>       WindowPtr[WindowHandle] = vme_master_window(BusHandle[deviceHandle]
>                               , WindowHandle[deviceHandle]...

The best would be if a mvme_read(...) to an unmapped region would automatically (internally) trigger a 
vme_map() call, and store the WindowHandle and WindowPtr internally. The advantage of this is that code 
written for the SIS for example (which does not require this kind of mapping) would work without change 
under the VMIC. The disadvantage is that for each mvme_read(), the code has to scan the internal mapping 
table to find the proper window handle. Now I don't know how much overhead this would be, but I guess a 
single for() loop over a couple of entries in the mapping table is still faster than a microsecond or so, 
thus making it negligible in a block transfer. 

> 3) vme_setmode(mapHandle, const int DATA_SIZE, const int AM
>                            , const BOOL ENA_DMA, const BOOL ENA_FIFO);
>     Mainly used for the vme_block_read/write. Define for following read the data size and 
>     am in case of DMA (could use orther DMA mode than window definition for optimal
>     transfer).
> 
>     Predefine the mode of access:
>     DATA_SIZE : D8, D16, D32
>     AM             : A16, A24, A32, etc...
>     enaDMA     : optional if available.
>     enaFIFO     : optional for block read for autoincrement source pointer.
> 
> Remark:
> PAA- I can imagine this function to be a vme_ioctl (int mapHandle, int *param)
>         such that extension of functionality is possible. But by passing cons int
>         arguments, the optimizer is able to substitute and reduce the internal code.

Right. mvme_ioctl(VME_IOCTL_AMOD_SET/DSIZE_SET/DMA_SET/AUTO_INCR_SET, ...)

>    uint_8Value   = vme_readD8  (int mapHandle, uint_64 vmeSrceOffset)
>    uint_16Value = vme_readD16 (int mapHandle, uint_64 vmeSrceOffset)
>    uint_32Value = vme_readD32 (int mapHandle, uint_64 vmeSrceOffset)
>    Single VME read access. In the VMIC case, this access is always through mapping.
>    Value = *(WindowPtr[WindowHandle] + vmeSrceOffset) 
>    or 
>    Value = *(WindowStruct->WindowPtr[WindowHandle] + vmeSrceOffset) 

mvme_read(*dst, DWORD vme_addr, DWORD size); would cover this in a single call. Note that the SIS for 
example does not have memory mapping, so if one consistently uses mvme_read(), it will work on both 
architectures. Again, this takes some overhead. Consider for example a possible VMIC implementation

mvme_read(char *dst, DWORD vme_addr, DWORD size)
{
  for (i=0 ; table[i].valid ; i++)
    {
    if (table[i].start >= vme_addr && table[i].end < vme_addr+size)
      break;
    }

  if (!table[i].valid)
    {
    vme_master_crate(...)
    table[i].window_handle = vme_master_window(...)
    }

  if (size == 2)
    mvme_ioctl(VME_IOCTL_DSIZE_SET, D16);
  else if (size == 1)
    mvme_ioctl(VME_IOCTL_DSIZE_SET, D8);

  memcpy(dst, table[i].window_handle + vme_addr - table[i].start, size);
}

Note this is only some rough code, would need more checking etc. But you see that for each access the for() 
loop has to be evaluated. Now I know that for the SBS/Bit3 and for the SIS a single VME access takes 
~0.5us. So the for() loop could be much faster than that. But one has to try. If one experiment needs the 
ultimate speed, it can use the native VMIC API, but then looses the portability. I'm not sure if one needs 
the automatic DSIZE_SET, maybe it works without.

>    status  = vme_writeD8   (int mapHandle, uint_64 vmeSrceOffset, uint_8 Value)
>    status  = vme_writeD16 (int mapHandle, uint_64 vmeSrceOffset, uint_16 Value)
>    status  = vme_writeD32 (int mapHandle, uint_64 vmeSrceOffset, uint_32 Value)
>    Single VME write access.

Dito. mvme_write(void *dst, DWORD vme_addr, DWORD size);

>    nBytes = vme_block_read(mapHandle, char * pDest, uint_64 vmeSrceOffset, int size);
>    Multiple read access. Can be done through standard do loop or DMA if available.
>    nBytes < 0 :  error
>    Incremented pDest  = (pDest + nBytes); Don't need to pass **pDest for autoincrement.

vmve_ioctl(VME_IOCTL_DMA_SET, TRUE);
n = mvme_read(char *pDest, DWORD vmd_addr, DWORD size);

>    nBytes = vme_block_write(mapHandle, uint_64 vmeSrceOffset, char *pSrce, int size);
>    Multiple write access.
>    nBytes < 0 :  error
>    Incremented pSrce  = (pSrce + nBytes); Don't need to pass **pSrce for autoincrement.

Dito.

> 8) status = vme_unmap(int mapHandle)
>    Cleanup internal pointers or structure of given mapHandle only.

mvme_unmap(DWORD vme_addr, DWORD size)

Scan through internal table to find handle, then calls vme_unmap(mapHandle);

> 9) status = vme_exit()
>    Cleanup deviceHandle and release device.

mvme_exit();

Let me know if this all makes sense to you...

- Stefan

Before we try to merge the different access scheme for the different VME hardware,
we present the "optimal" configuration for the VMIC setup. This is a first shot so take it
with caution.
From these definitions, we should be able to workout a compromise and come up with
a satisfactory standard.

A) The VMIC vme_slave_xxx() options are not considered.
B) The interrupt handling can certainly match the 4 entries required in the user frontend
    code i.e. Attach, Detach, Enable, Disable.

I don't understand your argument that the handle should be hidden. In case of multiple
interfaces, how do you refer to a particular one if not specified? 
The following scheme does require a handle for refering to the proper (device AND window).

1 ) deviceHandle = vme_init(int devNumber);
    Even though the VMIC doesn't deal with multiple devices,
    the SIS/PCI does and needs to init on a specific PCI card.
    Internally:
      opening of the device (/dev/sisxxxx_1) (ignored in case of VMIC).
      Possible including a mapping to a default VME region of default size with default AM
      (VMIC :16MB, A24). This way in a single call you get a valid handle for full VME access
      in A24 mode. Needs to be elaborate this option. But in principle you need to declare the 
     VME region that you want to work on (vme_map).

2) mapHandle = vme_map(int deviceHandle, int vmeAddress, int am, int size);
    Return a mapHandle specific to a device and window. The am has to be specified.
    What ever are the operation to get there, the mapHandle is a reference to thas setting.
    It could just fill a map structure.
    Internally:
      WindowHandle[deviceHandle] = vme_master_create(BusHandle[deviceHandle], ...
      WindowPtr[WindowHandle] = vme_master_window(BusHandle[deviceHandle]
                                                                           , WindowHandle[deviceHandle]...

3) vme_setmode(mapHandle, const int DATA_SIZE, const int AM
                           , const BOOL ENA_DMA, const BOOL ENA_FIFO);
    Mainly used for the vme_block_read/write. Define for following read the data size and 
    am in case of DMA (could use orther DMA mode than window definition for optimal
    transfer).

    Predefine the mode of access:
    DATA_SIZE : D8, D16, D32
    AM             : A16, A24, A32, etc...
    enaDMA     : optional if available.
    enaFIFO     : optional for block read for autoincrement source pointer.

Remark:
PAA- I can imagine this function to be a vme_ioctl (int mapHandle, int *param)
        such that extension of functionality is possible. But by passing cons int
        arguments, the optimizer is able to substitute and reduce the internal code.

4)   
   uint_8Value   = vme_readD8  (int mapHandle, uint_64 vmeSrceOffset)
   uint_16Value = vme_readD16 (int mapHandle, uint_64 vmeSrceOffset)
   uint_32Value = vme_readD32 (int mapHandle, uint_64 vmeSrceOffset)
   Single VME read access. In the VMIC case, this access is always through mapping.
   Value = *(WindowPtr[WindowHandle] + vmeSrceOffset) 
   or 
   Value = *(WindowStruct->WindowPtr[WindowHandle] + vmeSrceOffset) 
 
5)   
   status  = vme_writeD8   (int mapHandle, uint_64 vmeSrceOffset, uint_8 Value)
   status  = vme_writeD16 (int mapHandle, uint_64 vmeSrceOffset, uint_16 Value)
   status  = vme_writeD32 (int mapHandle, uint_64 vmeSrceOffset, uint_32 Value)
   Single VME write access.

6)
   nBytes = vme_block_read(mapHandle, char * pDest, uint_64 vmeSrceOffset, int size);
   Multiple read access. Can be done through standard do loop or DMA if available.
   nBytes < 0 :  error
   Incremented pDest  = (pDest + nBytes); Don't need to pass **pDest for autoincrement.

7)
   nBytes = vme_block_write(mapHandle, uint_64 vmeSrceOffset, char *pSrce, int size);
   Multiple write access.
   nBytes < 0 :  error
   Incremented pSrce  = (pSrce + nBytes); Don't need to pass **pSrce for autoincrement.

8) status = vme_unmap(int mapHandle)
   Cleanup internal pointers or structure of given mapHandle only.

9) status = vme_exit()
   Cleanup deviceHandle and release device.

Indeed the ODB command "cleanup" really works. I have used it several
times with the TWIST DAQ and regularly with the BNMR/MUSR setups where
we have these stubborn clients (ie feepics) that do not want to shutdown
cleanly.  
But there is one problem with "cleanup". It has a hardwired timeout of
2 seconds.  This is a problem for tasks like lazylogger which set a timeout
of 60 seconds when moving the tape. So BEWARE, if you issue the "cleanup"
command, it might kill some clients who have setup their timeout to longer
than 2 seconds. 

I have asked Stefan to change this before. He said that, to be effective,
the timeout value used for "cleanup" has to be rather short. 
One possibility, would be to allow for a user entered "cleanup" timeout.
The default could stay at 2 seconds. 




> > Have you tried a "cleanup" in ODBEdit?
> 
> Nope. Will try next time...
> 

> INT == "int", wraparound in 1 month
> DWORD == "unsigned int", wraparound in 2 months
> 
> should we make it the 64-bit "long long" (or C98's "int64_t")?

Won't work on all supported compilers. The point is that DWORD wraps around in 
2 months, but the difference of two DWORDs is alywas positive, never negative 
like you had it. We only have to distinguish if the difference of the current 
time (im ms) minus the last_activity of a client is larget than the timeout, 
typically 10 seconds or so. If you have a wraparound on 32-bit DWORD, the 
difference is still ok. Like

current "time" : 0x0000 0100
last_activity:   0xFFFF FF00

then current_time - last_activity = 0x00000100 - 0xFFFFFF00 = 0x00000200 if 
calculated with 32-bit values.

> > 1) shutdown from mhttpd "programs" page -> "cannot shutdown client"
> Have you tried a "cleanup" in ODBEdit?

Nope. Will try next time...

> The "last_activity" is a 32-bit int, filled with milliseconds. So indeed it 
> wraps around after about one month.... change last_activity 
> from INT to DWORD. This way it's alway positive and the wraparound does not 
> hurt.

INT == "int", wraparound in 1 month
DWORD == "unsigned int", wraparound in 2 months

should we make it the 64-bit "long long" (or C98's "int64_t")?

K.O.

> dio- is required to be setuid-root to gain I/O permissions. I looked at it a
> few times, and it is probably safe, but I would like to get a second
> opinion. Stephan, can you should it to your local security geeks?
> 
> mhttpd- definitely unsafe. It has more buffer overflows than I can shake a
> stick at. Why is it suid-root anyway?
> 
> webpaw- what is it?!?

dio was written by Pierre. 

mhttpd and webpaw both are web servers. webpaw is used to display PAW 
pictures over the web. If you run these programs at a port <1024, and most 
people do run them at port 80 (at least at PSI), they need to be setuid-root. 
Unless you know a better way to do that...

> 1) shutdown from mhttpd "programs" page -> "cannot shutdown client"
> 2) "sh mhttpd" from odbedit -> 
>    [midas.c:5298:cm_shutdown] cannot connect to client mhttpd on host
>    midtig01.triumf.ca, port 32853
>    Client mhttpd not active
> 3) in odbedit: "cd /system/clients; rm xxxx"
>    refuses to delete the key

Have you tried a "cleanup" in ODBEdit?

The "last_activity" is a 32-bit int, filled with milliseconds. So indeed it 
wraps around after about one month. So if a all clients are stopped 
simultaneously the hard way (such that nobody's watchdog can clean any other 
client from the ODB), like with a power off, and you start the thing one 
month later, there might be a problem. I never tried that before. So next 
time to a cleanup. If that does not help, we should change last_activity 
from INT to DWORD. This way it's alway positive and the wraparound does not 
hurt.

I see that MIDAS installs several set-uid-root programs into /usr/local/bin.
In this age and time of evil computer hackers, this is not a good idea and
we should Do Something (TM) about it. Here is my risk assessment:

[olchansk@midtis06 midas]$ ls -l /usr/local/bin | grep wsr
-rwsr-sr-x    1 root     root        25811 Nov 20 09:27 dio
-rwsr-sr-x    1 root     root       344553 Nov 20 09:27 mhttpd
-rwsr-sr-x    1 root     root        70736 Nov 20 09:27 webpaw

dio- is required to be setuid-root to gain I/O permissions. I looked at it a
few times, and it is probably safe, but I would like to get a second
opinion. Stephan, can you should it to your local security geeks?

mhttpd- definitely unsafe. It has more buffer overflows than I can shake a
stick at. Why is it suid-root anyway?

webpaw- what is it?!?

K.O.

> While reviving midas on midtig01 after it was not used for a while ... 
> [local:tigress:S]/>scl -w
> Name                Host                Timeout    Last called
> mhttpd              midtig01.triumf.ca  10000      -2037131082

These clients cannot be deleted. I tried:
1) shutdown from mhttpd "programs" page -> "cannot shutdown client"
2) "sh mhttpd" from odbedit -> 
   [midas.c:5298:cm_shutdown] cannot connect to client mhttpd on host
   midtig01.triumf.ca, port 32853
   Client mhttpd not active
3) in odbedit: "cd /system/clients; rm xxxx"
   refuses to delete the key

Lacking any better ideas, I deleted them via brain surgery on the odb file:
1) stop everything
2) ipcrm the SYSV shared memory segment
3) odbedit -> save xxx.odb
4) xemacs xxx.odb, delete offending odb entries
5) rm .ODB.SHM
6) odbedit -> load xxx.odb
7) voila, bad clients gone, gone, gone.

K.O.

While reviving midas on midtig01 after it was not used for a while, we see
this. Notice negative "last called" numbers. Looks like a time_t wraparound
somewhere...

[local:tigress:S]/>scl -w
Name                Host                Timeout    Last called
mhttpd              midtig01.triumf.ca  10000      -2037131082
Logger              midtig01.triumf.ca  10000      -2037131166
Analyzer            midtig01.triumf.ca  10000      -2037131048
JACQ                midtig01.triumf.ca  10000      -2037131667
mhttpd1             midtig01.triumf.ca  10000      325
ODBEdit             midtig01.triumf.ca  10000      829

K.O.

I tried to reproduce the problem, but without success. So in case this happens 
again, one should debug the code im cm_watchdog() next to the line

/* decrement notify_count for open records and clear exclusive mode */
...

So if a killed client is removed from the ODB via the watchdog (or a "cleanup" 
is done in ODBEdit), the notify_count should be decreased and thus the "open 
records" should be closed.