Обсуждение: Re: [PATCHES] patches for 6.2.1p6
> > Hi hackers, > > I have old patches for version 6.2.1p6 which fix some problems and add > new features. Here is a short description of each patch file: > > > assert.patch > > adds a switch to turn on/off the assert checking if enabled at compile > time. You can now compile postgres with assert checking and disable it > at runtime in a production environment. > > async-unlisten.patch > > declares Async_Unlisten() external so that it can be called by user > modules. > > exec-limit.patch > > removes the #ifdef NOT_USED around ExecutorLimit(). It is used. > > exitpg.patch > > limits recursive calls to exitpg() preventing an infinite loop > if an error is found inside exitpg. > > libpgtcl-listen.patch > > Just a change from upper to lowercase of an sql command in libpgtcl, > totally harmless. > > new-locks.patch > > After long studying and many debugging sessions I have finally > understood how the low level locks work. > I have completely rewritten lock.c cleaning up the code and adding > better assert checking. I have also added some fields to the lock > and xid tags for better support of user locks. This patch includes > also a patch submitted by Bruce Momjian which changes the handling > of lock priorities. It can however be disabled if an option is set > in pg_options, see tprintf.patch (Bruce patch works by building > the queue in reverse priority order, my old patch kept the queue in > decreasing order and traversed it from the other side). > > pg-flush.patch > > removes an unnecessary flush in libpq reducing network traffic and > increasing performance. > > relname.patch > > an utility which returns the relname corresponding to a given oid. > Useful for debug messages (see vacum.patch). > > sequence.patch > > added a setval() function which enables othe owner of a sequence > to set its value without need to delete and recreate it. > > sinval.patch > > fixes a problem in SI cache which causes table overflow if some > backend is idle for a long time while other backends keep adding > entries. > It uses the new signal handling implemented in tprintf.patch. > I have also increacasesed the max number of backends from 32 to 64 and > the table size from 1000 to 5000. > > spin-lock.patch > > I'm not sure if this is really useful, but it seems stupid to have > a backend wasting cpu cycles in a busy loop while the process which > should release the lock is waiting for the cpu. So I added a call > to process_yield() if the spin lock can't obtained. > This has been implemented and tested only on Linux. I don't know if > other OS have process_yield(). If someone can check please do it. > > tprintf.patch > > adds functions and macros which implement a conditional trace package > with the ability to change flags and numeric options of running > backends at runtime. > Options/flags can be specified in the command line and/or read from > the file pg_options in the data directory. > Running backends can be forced to update their options from this file > by sending them a SIGHUP signal (this is the convention used by most > unix daemons so I changed the meaning of SIGHUP). > Options can be debugging flags used by the trace package or any other > numeric value used by the backend, for example the deadlock_timeout. > Having flags and options specified at runtime and changed while the > backends are running can greatly simplify the debugging and tuning > of the database. New options can be defined in utils/misc/trace.c and > include/utils/trace.h. As an example of the usage of this package > see lock.c and proc.c which make use of new runtime options. > > Old files using int flags or variables can be easily changed to > use the new package by substituting the old variable with a #define > like in the following example: > > /* int my_flag = 0; */ > #include "trace.h" > #define my_flag pg_options[OPT_MYFLAG] > > I have done it in postgres.c and some other files and now I can turn > on/off any single debug flag on the fly with a simple shell script. > I have removed the IpcConfigTip() from ipc.c, it should better be > described in the postgres manual instead of being printed on stderr. > > This patch provides also a new format of debugging messages which > are always in a single line with a timestamp and the backend pid: > > #timestamp #pid #message > 980127.17:52:14.173 [29271] StartTransactionCommand > 980127.17:52:14.174 [29271] ProcessUtility: drop table t; > 980127.17:52:14.186 [29271] SIIncNumEntries: table is 70% full > 980127.17:52:14.186 [29286] Async_NotifyHandler > 980127.17:52:14.186 [29286] Waking up sleeping backend process > 980127.19:52:14.292 [29286] Async_NotifyFrontEnd > 980127.19:52:14.413 [29286] Async_NotifyFrontEnd done > 980127.19:52:14.466 [29286] Async_NotifyHandler done > > This improves the readability of the log and allows one to understand > exactly which backend is doing what and at which time. It also makes > easier to write simple awk or perl scripts which monitor the log to > detect database errors or problem, or to compute transaction times. > > The patch changes also the meaning of signals used by postgres, as > described by the following table: > > postmaster backend > > SIGHUP kill(*,sighup) read_pg_options > SIGINT kill(*,sigint), die die > SIGCHLD reaper - > SIGTTIN ignored - > SIGTTOU ignored - > SIGQUIT die handle_warn > SIGTERM kill(*,sigterm), kill(*,9), die die > SIGCONT dumpstatus - > SIGPIPE ignored die > SIGFPE - FloatExceptionHandler > SIGTSTP - ignored (alive test) > SIGUSR1 kill(*,sigusr1), die quickdie > SIGUSR2 kill(*,sigusr2) Async_NotifyHandler > (also SI buffer flush) > > The main changes to the old implementation are SIGQUIT instead of > SIGHUP to handle warns, SIGHUP to reread pg_options and redirection > to all backends of SIGHUP, SIGINT, SIGTERM, SIGUSR1 and SIGUSR2. > In this way some of the signals sent to the postmaster can be sent > automatically to all the backends. To shut down postgres one needs > only to send a SIGTERM to postmaster and it will stop automatically > all the backends. This new signal handling mechanism is also used > to prevent SI cache table overflows: when a backend detects the SI > table full at 70% it simply sends a signal to the postmaster which > will wake up all idle backends and make them flush the cache. > > vacuum.patch > > adds a debug message to vacuum that prints the name of a table or > index *before* vacuuming it, if the verbose keyword is set. > This is useful to know which table is causing troubles if a > vacuum all crashes. Currently table information is printed only > at the end of each vacuum operation and is never printed if the > vacuum crashes. > > -- > Massimo Dal Zotto Massimo, now that 6.3 is released, any chance of getting these patches against the 6.3 source code? -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
> > Hi hackers,
> >
> > I have old patches for version 6.2.1p6 which fix some problems and add
> > new features. Here is a short description of each patch file:
> > spin-lock.patch
> >
> > I'm not sure if this is really useful, but it seems stupid to have
> > a backend wasting cpu cycles in a busy loop while the process which
> > should release the lock is waiting for the cpu. So I added a call
> > to process_yield() if the spin lock can't obtained.
> > This has been implemented and tested only on Linux. I don't know if
> > other OS have process_yield(). If someone can check please do it.
The generic way to do this is
select( NULL_FDSET, NULL_FDSET, NULL_FDSET, &delaytime, NULL);
Delay time may be 0, but a random value between 0 and say 30 msec seems
to be optimal. Hard busy wait spinlocks cause huge performance problems with
heavily loaded systems and lots of postgres backends. Basically one backend
ends up with the lock and gets scheduled out holding it, every else queues
up busywaiting behind this one. But the backend holding the lock cannot
release it until all the other backeds waiting on the lock exhaust a full
timeslice busywaiting. Get 20 of these guys going (like on a busy website) and
the system pretty much stops doing any work at all.
I say we should get this in as soon as we can.
> > --
> > Massimo Dal Zotto
>
> Massimo, now that 6.3 is released, any chance of getting these patches
> against the 6.3 source code?
>
> --
> Bruce Momjian | 830 Blythe Avenue
> maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- I realize now that irony has no place in business communications.
On Mon, 16 Mar 1998, David Gould wrote:
> The generic way to do this is
>
> select( NULL_FDSET, NULL_FDSET, NULL_FDSET, &delaytime, NULL);
>
> Delay time may be 0, but a random value between 0 and say 30 msec seems
> to be optimal. Hard busy wait spinlocks cause huge performance problems with
> heavily loaded systems and lots of postgres backends. Basically one backend
> ends up with the lock and gets scheduled out holding it, every else queues
> up busywaiting behind this one. But the backend holding the lock cannot
> release it until all the other backeds waiting on the lock exhaust a full
> timeslice busywaiting. Get 20 of these guys going (like on a busy website) and
> the system pretty much stops doing any work at all.
>
> I say we should get this in as soon as we can.
Can you submit an appropriate patch that can be included in the mega-patch
to be created on Sunday?
Thanks...
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
>
> On Mon, 16 Mar 1998, David Gould wrote:
>
> > The generic way to do this is
> >
> > select( NULL_FDSET, NULL_FDSET, NULL_FDSET, &delaytime, NULL);
> >
> > Delay time may be 0, but a random value between 0 and say 30 msec seems
> > to be optimal. Hard busy wait spinlocks cause huge performance problems with
> > heavily loaded systems and lots of postgres backends. Basically one backend
> > ends up with the lock and gets scheduled out holding it, every else queues
> > up busywaiting behind this one. But the backend holding the lock cannot
> > release it until all the other backeds waiting on the lock exhaust a full
> > timeslice busywaiting. Get 20 of these guys going (like on a busy website) and
> > the system pretty much stops doing any work at all.
> >
> > I say we should get this in as soon as we can.
>
> Can you submit an appropriate patch that can be included in the mega-patch
> to be created on Sunday?
Just a warning that this is not going to be easy. We have OS-specific
code for spinlocks in include/storage/s_lock.h and
backend/storage/buffer/s_lock.c. So each S_LOCK macro call has to have
its test-and-set logic de-coupled with its while-lock-fail-try-again
logic. Most of them are easy, but some like VAX:
#define S_LOCK(addr) __asm__("1: bbssi $0,(%0),1b": :"r"(addr))
are hard to de-couple. Now, I did not know we supported NetBSD on VAX.
Does it work, anyone? Can I remove it?
This is going to be pretty tough to test on every platform we support,
so if it is done now, it will have to be done carefully.
--
Bruce Momjian | 830 Blythe Avenue
maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
+ If your life is a hard drive, | (610) 353-9879(w)
+ Christ can be your backup. | (610) 853-3000(h)
> > Can you submit an appropriate patch that can be included in the
> > mega-patch to be created on Sunday?
>
> Just a warning that this is not going to be easy. We have OS-specific
> code for spinlocks in include/storage/s_lock.h and
> backend/storage/buffer/s_lock.c. So each S_LOCK macro call has to
> have its test-and-set logic de-coupled with its
> while-lock-fail-try-again logic.
> Most of them are easy, but some like VAX:
>
> #define S_LOCK(addr) __asm__("1: bbssi $0,(%0),1b": :"r"(addr))
>
> are hard to de-couple. Now, I did not know we supported NetBSD on
> VAX. Does it work, anyone? Can I remove it?
NetBSD on VAX in on our supported list, and was verified for v6.3 by Tom
Helbekkmo.
> This is going to be pretty tough to test on every platform we support,
> so if it is done now, it will have to be done carefully.
Is this behavior in v6.2.x? In any case, if it is anything but minimally
trivial, it should be given a test on every supported platform, since it
hits the heart of the platform-specific code, doesn't it? Seems like it
should be put into the CVS tree and shaken out until the next release...
- Tom
On Tue, 17 Mar 1998, Thomas G. Lockhart wrote:
> > This is going to be pretty tough to test on every platform we support,
> > so if it is done now, it will have to be done carefully.
>
> Is this behavior in v6.2.x? In any case, if it is anything but minimally
> trivial, it should be given a test on every supported platform, since it
> hits the heart of the platform-specific code, doesn't it? Seems like it
> should be put into the CVS tree and shaken out until the next release...
Not realizing what was involved, I have to agree here...*after* I
get a post-release patch out on Sunday? :)
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
> > > > Can you submit an appropriate patch that can be included in the > > > mega-patch to be created on Sunday? > > are hard to de-couple. Now, I did not know we supported NetBSD on > > VAX. Does it work, anyone? Can I remove it? > > NetBSD on VAX in on our supported list, and was verified for v6.3 by Tom > Helbekkmo. > > > This is going to be pretty tough to test on every platform we support, > > so if it is done now, it will have to be done carefully. > > Is this behavior in v6.2.x? In any case, if it is anything but minimally > trivial, it should be given a test on every supported platform, since it > hits the heart of the platform-specific code, doesn't it? Seems like it > should be put into the CVS tree and shaken out until the next release... Yea, that is what I was hinting at. -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
> > > > Can you submit an appropriate patch that can be included in the > > > > mega-patch to be created on Sunday? > > > > are hard to de-couple. Now, I did not know we supported NetBSD on > > > VAX. Does it work, anyone? Can I remove it? > > > > NetBSD on VAX in on our supported list, and was verified for v6.3 by Tom > > Helbekkmo. > > > > > This is going to be pretty tough to test on every platform we support, > > > so if it is done now, it will have to be done carefully. > > > > Is this behavior in v6.2.x? In any case, if it is anything but minimally > > trivial, it should be given a test on every supported platform, since it > > hits the heart of the platform-specific code, doesn't it? Seems like it > > should be put into the CVS tree and shaken out until the next release... > > Yea, that is what I was hinting at. > > -- > Bruce Momjian | 830 Blythe Avenue > maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 I tend to agree but am willing to compromise. Can we do only the easy platforms at this time and then fix the others later? Since S_LOCK is a macro, it could be #define S_LOCK s_lock_with_backoff on the easy platforms and #define S_LOCK original_definition on the tricky or hard to test platforms If this will work, I am willing to hack this together tomorrow. What is the time frame for accepting a patch like this? -dg David Gould dg@illustra.com 510.628.3783 or 510.305.9468 Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612 - I realize now that irony has no place in business communications.
David Gould wrote:
>
> > > > > Can you submit an appropriate patch that can be included in the
> > > > > mega-patch to be created on Sunday?
> >
> > > > are hard to de-couple. Now, I did not know we supported NetBSD on
> > > > VAX. Does it work, anyone? Can I remove it?
> > >
> > > NetBSD on VAX in on our supported list, and was verified for v6.3 by Tom
> > > Helbekkmo.
> > >
> > > > This is going to be pretty tough to test on every platform we support,
> > > > so if it is done now, it will have to be done carefully.
> > >
> > > Is this behavior in v6.2.x? In any case, if it is anything but minimally
> > > trivial, it should be given a test on every supported platform, since it
> > > hits the heart of the platform-specific code, doesn't it? Seems like it
> > > should be put into the CVS tree and shaken out until the next release...
> >
> > Yea, that is what I was hinting at.
> >
> > --
> > Bruce Momjian | 830 Blythe Avenue
> > maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
>
> I tend to agree but am willing to compromise.
>
> Can we do only the easy platforms at this time
> If this will work, I am willing to hack this together tomorrow.
> What is the time frame for accepting a patch like this?
What do you think it would take to exercise the patch heavily enough on
any of the affected platforms to ensure that it behaves no worse than
the current code? Is a test case easy to set up and run? If so, you can
probably get ~5 platforms tested in the next few days, and reduce the
risk of including this in the mega-patch.
Alternatively, we could do this the day _after_ the mega-patch, so that
the two are decoupled, and have a nice "slock patch" a few days later.
At the moment, there aren't any large backend patches waiting to go (I
think Vadim is still sleeping to recover from v6.3 :)
- Tom
On Mon, 16 Mar 1998, David Gould wrote:
> If this will work, I am willing to hack this together tomorrow.
> What is the time frame for accepting a patch like this?
Assuming that its *clean* (clean meaning that Bruce fully approves
of it, as this is his area of the code...well, one of them
*grin*)...tomorrow would be great :) If Bruce has *any* doubts though, it
doesn't go in until after I do the patch I want to do...
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
> David Gould wrote: > > > > > > > > Can you submit an appropriate patch that can be included in the > > > > > > mega-patch to be created on Sunday? > > > > > > > > are hard to de-couple. Now, I did not know we supported NetBSD on > > > > > VAX. Does it work, anyone? Can I remove it? > > > > > > > > NetBSD on VAX in on our supported list, and was verified for v6.3 by Tom > > > > Helbekkmo. > > > > > > > > > This is going to be pretty tough to test on every platform we support, > > > > > so if it is done now, it will have to be done carefully. > > > > > > > > Is this behavior in v6.2.x? In any case, if it is anything but minimally > > > > trivial, it should be given a test on every supported platform, since it > > > > hits the heart of the platform-specific code, doesn't it? Seems like it > > > > should be put into the CVS tree and shaken out until the next release... > > > > > > Yea, that is what I was hinting at. > > > > > > -- > > > Bruce Momjian | 830 Blythe Avenue > > > maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 > > > > I tend to agree but am willing to compromise. > > > > Can we do only the easy platforms at this time > > If this will work, I am willing to hack this together tomorrow. > > What is the time frame for accepting a patch like this? > > What do you think it would take to exercise the patch heavily enough on > any of the affected platforms to ensure that it behaves no worse than > the current code? Is a test case easy to set up and run? If so, you can > probably get ~5 platforms tested in the next few days, and reduce the > risk of including this in the mega-patch. Simple but clumsy test: Multiple backends all doing single row inserts into private tables. Run it for a fixed time period and add up the total number of rows. Do this for both a small (eg 2) and large (eg 20 or 40) number of backends. psuedo perl /sh / sql follows # driver <number of users> <delay time> <run time> synchtime = timenow starttime = synctime + delaytime; endtime = starttime + runtime for user 1 to number of users singleuser synchtime delay_time run_time # singleuser <userno> <starttime> <endtime> exec sql 'create table user$user (i int);' sleep until start time n = 1; while timenow < endtime exec sql "insert into user$user values($n);" print "user $user inserted $n" The figure of merit is the total number of inserts. Oh and that the system does not fall over. But, it may be best to leave this until after the mega patch. I am not sure I want to share the blame ;-). Just to fill me in, where does the mega patch fall in with the next release snapshot? That is, if this misses the mega patch is it waiting until 6.4? -dg David Gould dg@illustra.com 510.628.3783 or 510.305.9468 Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612 - I realize now that irony has no place in business communications.
On Mon, 16 Mar 1998, David Gould wrote:
> Just to fill me in, where does the mega patch fall in with the next release
> snapshot? That is, if this misses the mega patch is it waiting until 6.4?
That is pretty much up to you, actually. There is nothing wrong
with a clean patch for this being placed in the patches directory, if it
can be done shortly after the post-release patch. Basically, starting
April 1st (or so), development basically starts up again, so backtracking
a patch from v6.4-alpha to v6.3 release can prove difficult :(
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
>
> On Mon, 16 Mar 1998, David Gould wrote:
>
> > If this will work, I am willing to hack this together tomorrow.
> > What is the time frame for accepting a patch like this?
>
> Assuming that its *clean* (clean meaning that Bruce fully approves
> of it, as this is his area of the code...well, one of them
> *grin*)...tomorrow would be great :) If Bruce has *any* doubts though, it
> doesn't go in until after I do the patch I want to do...
David, go for it. The code is all local in two files, and I think you
can basically change all the do{test-and-set} while(lock-is-false)
loops to:
do{test-and-set} while(lock-is-false && select ())
Pretty easy. No need to test multiple platforms. The ones where the
loop is integrated into the asm(), leave them for later.
--
Bruce Momjian | 830 Blythe Avenue
maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
+ If your life is a hard drive, | (610) 353-9879(w)
+ Christ can be your backup. | (610) 853-3000(h)
> But, it may be best to leave this until after the mega patch. I am not
> sure I want to share the blame ;-).
>
> Just to fill me in, where does the mega patch fall in with the next
> release snapshot? That is, if this misses the mega patch is it waiting
> until 6.4?
I would guess that we could post a separate patch any time soon,
especially since the changes are apparently isolated to only a few
places in the code. In the last release, we posted ~7 patches, each
independent of the others, and generated on our local source trees. Each
of the patches was, however, fairly simple, quite often only one or a
few lines of change, and were intended as bug fixes. Also, they were
easily tested. In any case, at a minimum the regression test should be
run (and passed!).
- Tom
* Bruce Momjian
|
| Just a warning that this is not going to be easy. We have OS-specific
| code for spinlocks in include/storage/s_lock.h and
| backend/storage/buffer/s_lock.c. So each S_LOCK macro call has to have
| its test-and-set logic de-coupled with its while-lock-fail-try-again
| logic. Most of them are easy, but some like VAX:
|
| #define S_LOCK(addr) __asm__("1: bbssi $0,(%0),1b": :"r"(addr))
|
| are hard to de-couple. Now, I did not know we supported NetBSD on VAX.
| Does it work, anyone? Can I remove it?
Yes, it works. No, please don't break it. Heck, I only just got it
in in time for 6.3! :-) The not-so-busy-waiting-spinlock stuff can be
put in on a platform at a time -- I'll expand the VAX version to do
the right thing once someone has done another platform, so I can see
what's the preferred way of doing it.
-tih
--
Popularity is the hallmark of mediocrity. --Niles Crane, "Frasier"
> * Bruce Momjian
> | Just a warning that this is not going to be easy. We have OS-specific
> | code for spinlocks in include/storage/s_lock.h and
> | backend/storage/buffer/s_lock.c. So each S_LOCK macro call has to have
> | its test-and-set logic de-coupled with its while-lock-fail-try-again
> | logic. Most of them are easy, but some like VAX:
> |
> | #define S_LOCK(addr) __asm__("1: bbssi $0,(%0),1b": :"r"(addr))
> |
> | are hard to de-couple. Now, I did not know we supported NetBSD on VAX.
> | Does it work, anyone? Can I remove it?
>
> Yes, it works. No, please don't break it. Heck, I only just got it
> in in time for 6.3! :-) The not-so-busy-waiting-spinlock stuff can be
> put in on a platform at a time -- I'll expand the VAX version to do
> the right thing once someone has done another platform, so I can see
> what's the preferred way of doing it.
>
> -tih
I won't. I hope.
Seriously, if you want to, please create a function to emulate the following:
/*
* tas(lock)
*
* Access to platform specific test_and_set functionality. Given pointer to
* lock attempts to acquire the lock atomically.
*
* Returns 0 for success, nonzero for failure.
*/
typedef slock_t unsigned char; /* or whatever works on the platform */
int tas(slock_t *lock)
{
slock_t tmp;
/* atomic, interlocked */
tmp = *lock;
*lock = -1; /* any nonzero will do here */
return (tmp != 0);
}
Given this, I can fold the VAX right into the grand scheme, just like a
normal computer (;-)).
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- I realize now that irony has no place in business communications.
>
> * Bruce Momjian
> |
> | Just a warning that this is not going to be easy. We have OS-specific
> | code for spinlocks in include/storage/s_lock.h and
> | backend/storage/buffer/s_lock.c. So each S_LOCK macro call has to have
> | its test-and-set logic de-coupled with its while-lock-fail-try-again
> | logic. Most of them are easy, but some like VAX:
> |
> | #define S_LOCK(addr) __asm__("1: bbssi $0,(%0),1b": :"r"(addr))
> |
> | are hard to de-couple. Now, I did not know we supported NetBSD on VAX.
> | Does it work, anyone? Can I remove it?
>
> Yes, it works. No, please don't break it. Heck, I only just got it
> in in time for 6.3! :-) The not-so-busy-waiting-spinlock stuff can be
> put in on a platform at a time -- I'll expand the VAX version to do
> the right thing once someone has done another platform, so I can see
> what's the preferred way of doing it.
OK, now I know that the VAX stuff is still used and supported. Good.
We don't have good platform-specific information on NetBSD and Linux
ports.
--
Bruce Momjian | 830 Blythe Avenue
maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
+ If your life is a hard drive, | (610) 353-9879(w)
+ Christ can be your backup. | (610) 853-3000(h)
> OK, now I know that the VAX stuff is still used and supported. Good.
> We don't have good platform-specific information on NetBSD and Linux
> ports.
??
Have you checked the hardcopy or html versions of the "Supported
Platforms" section in the Administrator's Guide? _Every_ entry in this
was updated and refreshed at the end of February.
What other info should we be collecting for this?
- Tom
> > > OK, now I know that the VAX stuff is still used and supported. Good. > > We don't have good platform-specific information on NetBSD and Linux > > ports. > > ?? > > Have you checked the hardcopy or html versions of the "Supported > Platforms" section in the Administrator's Guide? _Every_ entry in this > was updated and refreshed at the end of February. > > What other info should we be collecting for this? > > - Tom > Sorry, haven't looked there yet. -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
> > > > > Hi hackers, > > > > I have old patches for version 6.2.1p6 which fix some problems and add > > new features. Here is a short description of each patch file: > > > > ... > > > > -- > > Massimo Dal Zotto > > Massimo, now that 6.3 is released, any chance of getting these patches > against the 6.3 source code? I have already applied all my old patches against 6.3 except the lock patch. The lock code has changed and I have to verify if my patches are still compatible. It will take some time and I haven't very much. I found also an interesting bug in the notify code. I will post the new patches when I find some spare time to verify them. Could you please send me some documentation on the new lock and deadlock code in the meantime ? Massimo Dal Zotto +----------------------------------------------------------------------+ | Massimo Dal Zotto e-mail: dz@cs.unitn.it | | Via Marconi, 141 phone: ++39-461-534251 | | 38057 Pergine Valsugana (TN) www: http://www.cs.unitn.it/~dz/ | | Italy pgp: finger dz@tango.cs.unitn.it | +----------------------------------------------------------------------+
> I have already applied all my old patches against 6.3 except the lock
> patch. The lock code has changed and I have to verify if my patches are
Great.
> still compatible. It will take some time and I haven't very much.
> I found also an interesting bug in the notify code. I will post the new
> patches when I find some spare time to verify them.
> Could you please send me some documentation on the new lock and deadlock
> code in the meantime ?
Here is some comments from storage/lmgr/lock.c:DeadLockCheck(). This
was a real mind-bender for me. It finds holders of the lock it has been
waiting on, and checks to see if any of those holders is waiting on the
lock I own. If not, it then checks all the holders of locks these new
processes are waiting on, and checks to see if they are waiting on my
lock, and it continues until it has traced all backend process id's
related to my lock. I have a static checked_procs[] array that keeps
track of what I have checked so I don't loop. The code is recursive.
The new wait queue handling is documented in storage/lmgr/lock.c:ProcSleep().
---------------------------------------------------------------------------
* This code takes a list of locks a process holds, and the lock that
* the process is sleeping on, and tries to find if any of the processes
* waiting on its locks hold the lock it is waiting for. If no deadlock
* is found, it goes on to look at all the processes waiting on their locks.
*
* We have already locked the master lock before being called.
*/
bool
DeadLockCheck(SHM_QUEUE *lockQueue, LOCK *findlock, bool skip_check)
---------------------------------------------------------------------------
/*
* This is our only check to see if we found the lock we want.
*
* The lock we are waiting for is already in MyProc->lockQueue so we
* need to skip it here. We are trying to find it in someone
* else's lockQueue.
*/
if (lock == findlock && !skip_check)
---------------------------------------------------------------------------
/*
* For findlock's wait queue, we are interested in
* procs who are blocked waiting for a write-lock on
* the table we are waiting on, and already hold a
* lock on it. We first check to see if there is an
* escalation deadlock, where we hold a readlock and
* want a writelock, and someone else holds readlock
* on the same table, and wants a writelock.
*
* Basically, the test is, "Do we both hold some lock on
* findlock, and we are both waiting in the lock
* queue?"
*/
---------------------------------------------------------------------------
* For non-MyProc entries, we are looking only
* waiters, not necessarily people who already
* hold locks and are waiting. Now we check for
* cases where we have two or more tables in a
* deadlock. We do this by continuing to search
* for someone holding a lock
*/
if (DeadLockCheck(&(proc->lockQueue), findlock, false))
---------------------------------------------------------------------------
/*
* If the first entries in the waitQueue have a greater priority than
* we have, we must be a reader, and they must be a writers, and we
* must be here because the current holder is a writer or a reader but
* we don't share shared locks if a writer is waiting. We put
* ourselves after the writers. This way, we have a FIFO, but keep
* the readers together to give them decent priority, and no one
* starves. Because we group all readers together, a non-empty queue
* only has a few possible configurations:
*
* [readers] [writers] [readers][writers] [writers][readers]
* [writers][readers][writers]
*
* In a full queue, we would have a reader holding a lock, then a writer
* gets the lock, then a bunch of readers, made up of readers who
* could not share the first readlock because a writer was waiting,
* and new readers arriving while the writer had the lock.
*
*/
--
Bruce Momjian | 830 Blythe Avenue
maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
+ If your life is a hard drive, | (610) 353-9879(w)
+ Christ can be your backup. | (610) 853-3000(h)
>
> >
> > On Mon, 16 Mar 1998, David Gould wrote:
> >
> > > If this will work, I am willing to hack this together tomorrow.
> > > What is the time frame for accepting a patch like this?
> >
> > Assuming that its *clean* (clean meaning that Bruce fully approves
> > of it, as this is his area of the code...well, one of them
> > *grin*)...tomorrow would be great :) If Bruce has *any* doubts though, it
> > doesn't go in until after I do the patch I want to do...
>
> David, go for it. The code is all local in two files, and I think you
> can basically change all the do{test-and-set} while(lock-is-false)
> loops to:
>
> do{test-and-set} while(lock-is-false && select ())
>
> Pretty easy. No need to test multiple platforms. The ones where the
> loop is integrated into the asm(), leave them for later.
>
> --
> Bruce Momjian | 830 Blythe Avenue
> maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
> + If your life is a hard drive, | (610) 353-9879(w)
> + Christ can be your backup. | (610) 853-3000(h)
>
>
>
I'am against a generic patch using select(). If we have sched_yield() on an
architecture I don't see why dont't use it. Here is the patch for Linux.
It has been tested for two months by 100 users without any problem.
The only thing I would add is a more general configuration test in configure
to include the proper include files.
*** src/include/storage/s_lock.h.orig Sat Oct 18 22:39:21 1997
--- src/include/storage/s_lock.h Wed Nov 19 23:11:14 1997
***************
*** 294,300 ****
--- 294,314 ----
*/
#if defined(NEED_I386_TAS_ASM)
+ #include <unistd.h>
+ #include <sched.h>
+ #ifdef _POSIX_PRIORITY_SCHEDULING
+ #define S_LOCK(lock) do \
+ { \
+ slock_t _res; \
+ do \
+ { \
+ __asm__("xchgb %0,%1": "=q"(_res), \
+ "=m"(*lock):"0"(0x1)); \
+ if (_res) sched_yield(); \
+ } while (_res != 0); \
+ } while (0)
+ #else
#define S_LOCK(lock) do \
{ \
slock_t _res; \
***************
*** 303,308 ****
--- 317,323 ----
__asm__("xchgb %0,%1": "=q"(_res), "=m"(*lock):"0"(0x1)); \
} while (_res != 0); \
} while (0)
+ #endif
#define S_UNLOCK(lock) (*(lock) = 0)
Massimo Dal Zotto
+----------------------------------------------------------------------+
| Massimo Dal Zotto e-mail: dz@cs.unitn.it |
| Via Marconi, 141 phone: ++39-461-534251 |
| 38057 Pergine Valsugana (TN) www: http://www.cs.unitn.it/~dz/ |
| Italy pgp: finger dz@tango.cs.unitn.it |
+----------------------------------------------------------------------+
> I'am against a generic patch using select(). If we have sched_yield() on an > architecture I don't see why dont't use it. Here is the patch for Linux. > It has been tested for two months by 100 users without any problem. > The only thing I would add is a more general configuration test in configure > to include the proper include files. I understand your issue. Unfortunately, only Linux has sched_yield(), as far as I know. Perhaps we can implement sched_yield/select based on the platform. However, if someone is holding a spinlock, does sched_yield() give the other process enough time to finish with the spinlock before we start checking it again. Seems select() allows us to control the time we wait before checking again. Also, it looks like the s_lock.h file is going to change pretty radically from David's change, so when he is done, we can put some OS-specific stuff if you wish. -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
Massimo Dal Zotto writes:
> > David, go for it. The code is all local in two files, and I think you
> > can basically change all the do{test-and-set} while(lock-is-false)
> > loops to:
> >
> > do{test-and-set} while(lock-is-false && select ())
> >
> > Pretty easy. No need to test multiple platforms. The ones where the
> > loop is integrated into the asm(), leave them for later.
> >
> > --
> > Bruce Momjian | 830 Blythe Avenue
>
> I'am against a generic patch using select(). If we have sched_yield() on an
> architecture I don't see why dont't use it. Here is the patch for Linux.
> It has been tested for two months by 100 users without any problem.
> The only thing I would add is a more general configuration test in configure
> to include the proper include files.
>
> *** src/include/storage/s_lock.h.orig Sat Oct 18 22:39:21 1997
> --- src/include/storage/s_lock.h Wed Nov 19 23:11:14 1997
> ***************
> *** 294,300 ****
> --- 294,314 ----
> */
>
> #if defined(NEED_I386_TAS_ASM)
> + #include <unistd.h>
> + #include <sched.h>
>
> + #ifdef _POSIX_PRIORITY_SCHEDULING
> + #define S_LOCK(lock) do \
> + { \
> + slock_t _res; \
> + do \
> + { \
> + __asm__("xchgb %0,%1": "=q"(_res), \
> + "=m"(*lock):"0"(0x1)); \
> + if (_res) sched_yield(); \
> + } while (_res != 0); \
> + } while (0)
> + #else
> #define S_LOCK(lock) do \
> { \
> slock_t _res; \
> ***************
> *** 303,308 ****
> --- 317,323 ----
> __asm__("xchgb %0,%1": "=q"(_res), "=m"(*lock):"0"(0x1)); \
> } while (_res != 0); \
> } while (0)
> + #endif
>
> #define S_UNLOCK(lock) (*(lock) = 0)
>
> +----------------------------------------------------------------------+
> | Massimo Dal Zotto e-mail: dz@cs.unitn.it |
> | Via Marconi, 141 phone: ++39-461-534251 |
> | 38057 Pergine Valsugana (TN) www: http://www.cs.unitn.it/~dz/ |
> | Italy pgp: finger dz@tango.cs.unitn.it |
> +----------------------------------------------------------------------+
I am perfectly happy to use sched_yield() on Linux. My goal however is to
make the concept work on all platforms.
There was a recent thread on comp.os.linux.system on context switch times
of Linux vs NTthat revealed that sched_yield() is fairly costly as it causes
the scheduler to do a full scan of the process table and recalculate the
the priorities of all processes. Probably not a problem, but it should
be it should probably be benchmarked both ways.
Finally, even though this appears to work, there is a possible stability
problem with both approaches. Here is some of the discussion I had about
that with Bruce:
-----------------------------------------------------------------------
David:
> Right now, there is not much chance of catching a signal while waiting for
> a spinlock. This is good, cause the code that waits for spinlocks tends to
> be doing funny things with buffers and locks and shared stuff like that.
> We don't catch signals because we don't make syscalls. But, once this goes in,
> we will be calling select() a _lot_ and so it kinda becomes the likely place
> for a signal to get delivered. Without thinking about it and looking at the
> code a bit longer, I am not sure it is prudent to rush this in. I still want
> it in as soon a possible, but possible includes free from harmful side effects.
Bruce:
> Well, signals are handled in the backend by tcop/postgres.c. In
> most/all? cases, a signal causes a longjump() out of the code and
> releases locks and aborts the transaction.
David:
> I was afraid that would be the answer. Basically, this never worked. The
> problem is that in an indeterminate number of places the code manipulates
> multiple distinct but related structures in a sequence. To leave the server
> in a consistant state all these updates need to be done in the right order
> so that if the sequence in interrupted by a signal the partially updated
> state is consistant. Unhappily, the original coders did not always have this
> in mind. An example:
>
> cleaning up after a scan
> 1 - release buffer pointed by scan descriptor
> 2 - release scan descriptor
>
> If a signal is taken just after one, the abort code will see the scan
> descriptor and call the cleanup for it resulting in:
>
> cleaning up after a scan (take 2)
> 1 - release buffer pointed by scan descriptor
> - Whoopsie, buffer already released!
> 2 - release scan descriptor
>
> These sequences either _all_ have to identified and fixed, or made atomic
> somehow, which is a biggish job.
>
> Or the system has to acknowledge signals at only clearly defined points.
> My preference is to have signal handlers only set a global flag to indicate
> that a signal was seen. Then one just sprinkles tests of the flag in
> all the likely places: step to next node, fetch next page, call function, etc.
>
> The way this shows up in real life is strange unreproduceable errors on busy
> systems, especially when backends are killed or transactions are forced to
> abort.
>
> Fixing this is a bit of a project, but the result is that a lot of mystery
> bugs go away.
--------------------------------------------------------------------------
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- I realize now that irony has no place in business communications.
On Fri, 20 Mar 1998, Bruce Momjian wrote:
> > I'am against a generic patch using select(). If we have sched_yield() on an
> > architecture I don't see why dont't use it. Here is the patch for Linux.
> > It has been tested for two months by 100 users without any problem.
> > The only thing I would add is a more general configuration test in configure
> > to include the proper include files.
>
> I understand your issue. Unfortunately, only Linux has sched_yield(),
> as far as I know. Perhaps we can implement sched_yield/select based on
> the platform.
What's the possibility of doing this similar to how we do some of
the other functions (dl_open comes immediately to mind)...make a
pg_sched_yield function and use that, which is built based on the various
platforms?
Right now, I don't believe we have *anything* in place, so have
pg_sched_yield() return 0 (or an equivalent) for every platform except for
Linux...
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
>
> On Fri, 20 Mar 1998, Bruce Momjian wrote:
>
> > > I'am against a generic patch using select(). If we have sched_yield() on an
> > > architecture I don't see why dont't use it. Here is the patch for Linux.
> > > It has been tested for two months by 100 users without any problem.
> > > The only thing I would add is a more general configuration test in configure
> > > to include the proper include files.
> >
> > I understand your issue. Unfortunately, only Linux has sched_yield(),
> > as far as I know. Perhaps we can implement sched_yield/select based on
> > the platform.
>
> What's the possibility of doing this similar to how we do some of
> the other functions (dl_open comes immediately to mind)...make a
> pg_sched_yield function and use that, which is built based on the various
> platforms?
>
> Right now, I don't believe we have *anything* in place, so have
> pg_sched_yield() return 0 (or an equivalent) for every platform except for
> Linux...
>
> Marc G. Fournier
> Systems Administrator @ hub.org
> primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
I appreciate all the help, but I think I have a solution for this. Details
next week...
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- I realize now that irony has no place in business communications.
> What's the possibility of doing this similar to how we do some of > the other functions (dl_open comes immediately to mind)...make a > pg_sched_yield function and use that, which is built based on the various > platforms? > > Right now, I don't believe we have *anything* in place, so have > pg_sched_yield() return 0 (or an equivalent) for every platform except for > Linux... > Probably even easier. Just use #ifdef linux around the select or sched_yield. -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
The Hermit Hacker wrote: > > What's the possibility of doing this similar to how we do some of > the other functions (dl_open comes immediately to mind)...make a > pg_sched_yield function and use that, which is built based on the various > platforms? > > Right now, I don't believe we have *anything* in place, so have > pg_sched_yield() return 0 (or an equivalent) for every platform except for > Linux... But sched_yield() is not Linux-specific: -- The sched_yield() function relinquishes the processor for the -- running process. -- IEEE Std 1003.1b-1993, §13.3.5. (POSIX real-time standard 1003.lb) Except from Linux, I can find references to sched_yield() in LynxOS, DECthreads thread library, AIX 4.1 and up (libc), Solaris (thread.h (c)1994 Sun Microsystems), Unix98, GNU, C EXECUTIVE(r) and PSX(tm) real time kernels ... This is just a quick search. Perhaps we should enable sched_yield() for every OS except for... well, what's the name of that OS which does not have sched_yield()... FreeBSD ;) After all, sched_yield() is five years old. Any reasonable OS should have it. /* m */
On Sun, 22 Mar 1998, Mattias Kregert wrote:
> The Hermit Hacker wrote:
> >
> > What's the possibility of doing this similar to how we do some of
> > the other functions (dl_open comes immediately to mind)...make a
> > pg_sched_yield function and use that, which is built based on the various
> > platforms?
> >
> > Right now, I don't believe we have *anything* in place, so have
> > pg_sched_yield() return 0 (or an equivalent) for every platform except for
> > Linux...
>
> But sched_yield() is not Linux-specific:
> -- The sched_yield() function relinquishes the processor for the
> -- running process.
> -- IEEE Std 1003.1b-1993, �13.3.5. (POSIX real-time standard 1003.lb)
>
> Except from Linux, I can find references to sched_yield() in LynxOS,
> DECthreads thread library, AIX 4.1 and up (libc), Solaris (thread.h
> (c)1994 Sun
> Microsystems), Unix98, GNU, C EXECUTIVE(r) and PSX(tm) real time kernels
> ...
> This is just a quick search.
>
> Perhaps we should enable sched_yield() for every OS except for... well,
> what's the
> name of that OS which does not have sched_yield()... FreeBSD ;)
>
> After all, sched_yield() is five years old. Any reasonable OS should
> have it.
You missed my point...so far as I've heard, there are two ways of
doing what is being proposed...either using sched_yield() on those systems
that support it, or select() on those that don't. If you are going to
build a patch for this, it should look something like:
#ifdef HAVE_SCHED_YIELD
<insert sched_yield() code here>
#else
<insert select() code here>
#endif
Totally 'configure' configurable, and non-system dependent :)
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
> On Sun, 22 Mar 1998, Mattias Kregert wrote:
>
> > The Hermit Hacker wrote:
> > >
> > > What's the possibility of doing this similar to how we do some of
> > > the other functions (dl_open comes immediately to mind)...make a
> > > pg_sched_yield function and use that, which is built based on the various
> > > platforms?
> > >
> > > Right now, I don't believe we have *anything* in place, so have
> > > pg_sched_yield() return 0 (or an equivalent) for every platform except for
> > > Linux...
> >
> > But sched_yield() is not Linux-specific:
> > -- The sched_yield() function relinquishes the processor for the
> > -- running process.
> > -- IEEE Std 1003.1b-1993, �13.3.5. (POSIX real-time standard 1003.lb)
> >
> > Except from Linux, I can find references to sched_yield() in LynxOS,
> > DECthreads thread library, AIX 4.1 and up (libc), Solaris (thread.h
> > (c)1994 Sun
> > Microsystems), Unix98, GNU, C EXECUTIVE(r) and PSX(tm) real time kernels
> > ...
> > This is just a quick search.
> >
> > Perhaps we should enable sched_yield() for every OS except for... well,
> > what's the
> > name of that OS which does not have sched_yield()... FreeBSD ;)
> >
> > After all, sched_yield() is five years old. Any reasonable OS should
> > have it.
>
> You missed my point...so far as I've heard, there are two ways of
> doing what is being proposed...either using sched_yield() on those systems
> that support it, or select() on those that don't. If you are going to
> build a patch for this, it should look something like:
>
> #ifdef HAVE_SCHED_YIELD
> <insert sched_yield() code here>
> #else
> <insert select() code here>
> #endif
>
> Totally 'configure' configurable, and non-system dependent :)
>
> Marc G. Fournier
> Systems Administrator @ hub.org
> primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
Ok, I will add a configuration option
USE_SCHED_YIELD
to the select patch I am working on. This can be enabled by configure.
Assuming someone can find the header files needed on all the platforms.
However, we should not assume that sched_yield() even where available is
the automatic "right thing". It might be, but...
The situation that either the select() or the sched_yield() style of
spinlock back off is meant to help is when there are a number of processes
busywaiting on the same spinlock.
On Linux, sched_yield() triggers the scheduler to to a full priority re-calc
for all processses. This is slightly expensive especially with a long run
queue. Having a bunch of processes pound on sched_yield() might be actually
worse than to use select(). At the very least it needs testing.
Secondly, the select() backoff patch I am working on starts out with a zero
timeout and backs off incrementally by increasing the timeout value on
subsequent iterations. The idea is to break up convoys and avoid big piles of
processes pounding on a spinlock. This cannot be done with sched_yield().
Which is better? Well, golly gosh, I have no idea. I know that the select()
flavor effectively solves the problem caused by hard loop busy waiting.
Without some testing it is kinda hard to say more than that. I will try to
fit in some testing, but if someone has a favorite many process workload
and would like to try comparing both flavors it would be useful.
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- I realize now that irony has no place in business communications.
> > You missed my point...so far as I've heard, there are two ways of > doing what is being proposed...either using sched_yield() on those systems > that support it, or select() on those that don't. If you are going to > build a patch for this, it should look something like: > > #ifdef HAVE_SCHED_YIELD > <insert sched_yield() code here> > #else > <insert select() code here> > #endif > > Totally 'configure' configurable, and non-system dependent :) Yep, I like it. -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
On Sat, 21 Mar 1998, David Gould wrote:
> Ok, I will add a configuration option
>
> USE_SCHED_YIELD
>
> to the select patch I am working on. This can be enabled by configure.
> Assuming someone can find the header files needed on all the platforms.
>
> However, we should not assume that sched_yield() even where available is
> the automatic "right thing". It might be, but...
>
> The situation that either the select() or the sched_yield() style of
> spinlock back off is meant to help is when there are a number of processes
> busywaiting on the same spinlock.
>
> On Linux, sched_yield() triggers the scheduler to to a full priority re-calc
> for all processses. This is slightly expensive especially with a long run
> queue. Having a bunch of processes pound on sched_yield() might be actually
> worse than to use select(). At the very least it needs testing.
>
> Secondly, the select() backoff patch I am working on starts out with a zero
> timeout and backs off incrementally by increasing the timeout value on
> subsequent iterations. The idea is to break up convoys and avoid big piles of
> processes pounding on a spinlock. This cannot be done with sched_yield().
>
> Which is better? Well, golly gosh, I have no idea. I know that the select()
> flavor effectively solves the problem caused by hard loop busy waiting.
> Without some testing it is kinda hard to say more than that. I will try to
> fit in some testing, but if someone has a favorite many process workload
> and would like to try comparing both flavors it would be useful.
Okay, we have two differing viewpoints here...from what I've been
able to read, the select() solution will work on *all* platforms, while
the sched_yield() will work on *some* systems, but not all.
I personally like the "work on all platform" solution, but that's
just me :)
I may have missed it, but I'm curious as to under what
circumstance sched_yield() is better then the select() solution? The
"con", as I see it, is that sched_yield() doesn't work everywhere...
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
> Perhaps we should enable sched_yield() for every OS except for... well,
> what's the
> name of that OS which does not have sched_yield()... FreeBSD ;)
>
> After all, sched_yield() is five years old. Any reasonable OS should
> have it.
Gee, I just checked and BSDI has it. However, it appears to work only
on threaded applications:
#include <pthread.h>
...
If other threads are ready to run, the sched_yield() function forces the
current thread to suspend itself temporarily and let them execute.
--
Bruce Momjian | 830 Blythe Avenue
maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026
+ If your life is a hard drive, | (610) 353-9879(w)
+ Christ can be your backup. | (610) 853-3000(h)
> Secondly, the select() backoff patch I am working on starts out with a zero > timeout and backs off incrementally by increasing the timeout value on > subsequent iterations. The idea is to break up convoys and avoid big piles of > processes pounding on a spinlock. This cannot be done with sched_yield(). Hard to beat the backoff argument. I vote we only use select(). -- Bruce Momjian | 830 Blythe Avenue maillist@candle.pha.pa.us | Drexel Hill, Pennsylvania 19026 + If your life is a hard drive, | (610) 353-9879(w) + Christ can be your backup. | (610) 853-3000(h)
On Sat, 21 Mar 1998, Bruce Momjian wrote:
> > Secondly, the select() backoff patch I am working on starts out with a zero
> > timeout and backs off incrementally by increasing the timeout value on
> > subsequent iterations. The idea is to break up convoys and avoid big piles of
> > processes pounding on a spinlock. This cannot be done with sched_yield().
>
> Hard to beat the backoff argument. I vote we only use select().
I haven't heard any compelling arguments so far as to why
sched_yield() is better then select(), so I tend to vote the same way...
Marc G. Fournier
Systems Administrator @ hub.org
primary: scrappy@hub.org secondary: scrappy@{freebsd|postgresql}.org
On Sun, 22 Mar 1998, Mattias Kregert wrote: > Except from Linux, I can find references to sched_yield() in LynxOS, > DECthreads thread library, AIX 4.1 and up (libc), Solaris (thread.h > (c)1994 Sun > Microsystems), Unix98, GNU, C EXECUTIVE(r) and PSX(tm) real time kernels > ... > This is just a quick search. This seems to be part of Posix threads, which means that sched_yield should only be callable by a thread.... I'm surprised that normal processes can call sched_yield. In fact, in most of the environments you've mentioned, it certainly can't be used outside a thread. Tom
David Gould wrote:
> Seriously, if you want to, please create a function to emulate the following:
>
> /*
> * tas(lock)
> *
> * Access to platform specific test_and_set functionality. Given pointer to
> * lock attempts to acquire the lock atomically.
> *
> * Returns 0 for success, nonzero for failure.
> */
> typedef slock_t unsigned char; /* or whatever works on the platform */
>
> int tas(slock_t *lock)
> {
> slock_t tmp;
>
> /* atomic, interlocked */
> tmp = *lock;
> *lock = -1; /* any nonzero will do here */
>
> return (tmp != 0);
> }
>
> Given this, I can fold the VAX right into the grand scheme, just like a
> normal computer (;-)).
Hmpf! The true worth of a computer is a function of its weight! :-)
Sorry this took a while, but anyway, this should do it for the VAX (in
fact, it's more or less the version of the code that I figured I'd use
until Bruce asked me to bum it down maximally for performance, only
now with the return values from tas() swapped). I include the macros
that would fit the current (6.3) locking scheme:
typedef unsigned char slock_t;
int tas(slock_t *lock) {
register ret;
asm(" movl $1, r0
bbssi $0,(%1),1f
clrl r0
1: movl r0,%0"
: "=r"(ret) /* return value, in register */
: "r"(lock) /* argument, 'lock pointer', in register */
: "r0"); /* inline code uses this register */
return ret;
}
#define S_LOCK(addr) do { while (tas(addr)) ; } while (0)
#define S_UNLOCK(addr) (*(addr) = 0)
#define S_INIT_LOCK(addr) (*(addr) = 0)
-tih
--
Popularity is the hallmark of mediocrity. --Niles Crane, "Frasier"
> David Gould wrote:
>
> >Seriously, if you want to, please create a function to emulate the following:
> >
> > /*
> > * tas(lock)
> > *
> > * Access to platform specific test_and_set functionality. Given pointer to
> > * lock attempts to acquire the lock atomically.
> > *
> > * Returns 0 for success, nonzero for failure.
> > */
> > typedef slock_t unsigned char; /* or whatever works on the platform */
> >
> > int tas(slock_t *lock)
> > {
> > slock_t tmp;
> >
> > /* atomic, interlocked */
> > tmp = *lock;
> > *lock = -1; /* any nonzero will do here */
> >
> > return (tmp != 0);
> > }
> >
> > Given this, I can fold the VAX right into the grand scheme, just like a
> > normal computer (;-)).
>
> Hmpf! The true worth of a computer is a function of its weight! :-)
>
> Sorry this took a while, but anyway, this should do it for the VAX (in
> fact, it's more or less the version of the code that I figured I'd use
> until Bruce asked me to bum it down maximally for performance, only
> now with the return values from tas() swapped). I include the macros
What do you mean "now with the return values from tas() swapped"? I think
your code looks ok, but just want to be sure we are following the same
grand plan...
> that would fit the current (6.3) locking scheme:
>
> typedef unsigned char slock_t;
>
> int tas(slock_t *lock) {
> register ret;
>
> asm(" movl $1, r0
> bbssi $0,(%1),1f
> clrl r0
> 1: movl r0,%0"
> : "=r"(ret) /* return value, in register */
> : "r"(lock) /* argument, 'lock pointer', in register */
> : "r0"); /* inline code uses this register */
>
> return ret;
> }
>
> #define S_LOCK(addr) do { while (tas(addr)) ; } while (0)
> #define S_UNLOCK(addr) (*(addr) = 0)
> #define S_INIT_LOCK(addr) (*(addr) = 0)
>
> -tih
> --
Thanks, this is just what I was looking for. I will fold it in to my changes.
I have gotten a little snowed under with other tasks, but I expect to finalize
my patch next week and will post it.
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- Linux. Not because it is free. Because it is better.
* David Gould
|
| What do you mean "now with the return values from tas() swapped"? I think
| your code looks ok, but just want to be sure we are following the same
| grand plan...
I just meant that my original code (which has been posted before) had
the tas() function implemented so that it returned 0 on failure, not
on success, as you asked for. Thus, I had to swap the sense of the
return value. In practice, I changed
clrl r0 ; clear register r0
bbssi $0,(%1),1f ; branch on bit set else set
incl r0 ; increment register r0
1: movl r0,%0 ; return register r0
[...]
#define S_LOCK(addr) do { while (!tas(addr)) ; } while (0)
...into...
movl $1, r0 ; set register r0 to 1
bbssi $0,(%1),1f ; branch on bit set else set
clrl r0 ; clear register r0
1: movl r0,%0 ; return register r0
[...]
#define S_LOCK(addr) do { while (tas(addr)) ; } while (0)
-tih
--
Popularity is the hallmark of mediocrity. --Niles Crane, "Frasier"
Tom Ivar Helbekkmo:
> * David Gould
> |
> | What do you mean "now with the return values from tas() swapped"? I think
> | your code looks ok, but just want to be sure we are following the same
> | grand plan...
>
> I just meant that my original code (which has been posted before) had
> the tas() function implemented so that it returned 0 on failure, not
> on success, as you asked for. Thus, I had to swap the sense of the
> return value. In practice, I changed
>
> clrl r0 ; clear register r0
> bbssi $0,(%1),1f ; branch on bit set else set
> incl r0 ; increment register r0
> 1: movl r0,%0 ; return register r0
>
> [...]
>
> #define S_LOCK(addr) do { while (!tas(addr)) ; } while (0)
>
> ...into...
>
> movl $1, r0 ; set register r0 to 1
> bbssi $0,(%1),1f ; branch on bit set else set
> clrl r0 ; clear register r0
> 1: movl r0,%0 ; return register r0
>
> [...]
>
> #define S_LOCK(addr) do { while (tas(addr)) ; } while (0)
>
Thats what I thought, but it has been a few years (not saying how many ;-) )
since I wrote any Macro-32 so I figured I should check.
The tas() definition is not to return success or failure so much as to
return the _previous_state_ of the lock. So if you test_and_set the lock
the test part returns true if it was previously locked and false if it was
unlocked. In either case, it is locked after the tas() (the set part).
Only, if it was previously unlocked, someone else owns the lock so we
have to wait for them to unlock it.
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
- Linux. Not because it is free. Because it is better.