[PATCH i-g-t v8 3/8] lib/drmtest: allow opening cards out of order

Kamil Konieczny kamil.konieczny at linux.intel.com
Fri Feb 9 19:29:17 UTC 2024


Hi Janusz,
On 2024-02-09 at 18:02:37 +0100, Janusz Krzysztofik wrote:
> Hi Kamil,
> 
> On Thursday, 8 February 2024 14:52:27 CET Kamil Konieczny wrote:
> > Current __drm_open_driver_another() implementation prevents
> > opening cards out of order, for example card 3 without previous
> > opens of 0..2. This can be seen with three GPU cards system and
> > gem_open_basic test:
> > 
> > sudo IGT_DEVICE=pci:vendor=Intel,card=all build/tests/gem_basic --r multigpu-create-close
> > 
> > IGT-Version: 1.28-g2548a539e (x86_64) (Linux: 6.7.0-rc5 x86_64)
> > Using IGT_SRANDOM=1706630533 for randomisation
> > Opened device: /dev/dri/card0
> > Starting subtest: multigpu-create-close
> > <g:0> Testing creating and closing an object.
> > <g:1> Opened device: /dev/dri/card1
> > <g:1> Testing creating and closing an object.
> > gem_basic: ../lib/drmtest.c:341: _is_already_opened: Assertion `as_idx <= _opened_fds_count' failed.
> > Received signal SIGABRT.
> > 
> > Relax that condition and add some more debug prints.
> > 
> > Changes in __search_and_open() was suggested by Zbigniew.
> > 
> > v5: removed caching for names of opened cards
> >  relaxed opening of filtered devices
> > v6: fixed assingment in _is_already_open() (Janusz)
> >  simplify use of this function
> > v7: restore check for already opened filtered card (Janusz)
> > v8: limit checks for N-th card to 0...N-1 range (Janusz)
> 
> As we discussed, that was not my intention.
> 

It may be reverted, see my proposition below.

> >  simplify if-else in __search_and_open() (Kamil)
> > 
> > Cc: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> > Cc: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > Signed-off-by: Kamil Konieczny <kamil.konieczny at linux.intel.com>
> > ---
> >  lib/drmtest.c | 73 +++++++++++++++++++++++++++++++++++++++------------
> >  1 file changed, 56 insertions(+), 17 deletions(-)
> > 
> > diff --git a/lib/drmtest.c b/lib/drmtest.c
> > index 73d9159af..62d31b000 100644
> > --- a/lib/drmtest.c
> > +++ b/lib/drmtest.c
> > @@ -322,35 +322,54 @@ static int _opened_fds_count;
> >  static void _set_opened_fd(int idx, int fd)
> >  {
> >  	assert(idx < ARRAY_SIZE(_opened_fds));
> > -	assert(idx <= _opened_fds_count);
> > +	assert(idx >= 0);
> >  
> >  	_opened_fds[idx].fd = fd;
> >  
> >  	assert(fstat(fd, &_opened_fds[idx].stat) == 0);
> >  
> > -	_opened_fds_count = idx+1;
> > +	for (int n = _opened_fds_count; n < idx; n++)
> > +		_opened_fds[n].fd = -1;
> > +
> > +	if (idx >= _opened_fds_count)
> > +		_opened_fds_count = idx + 1;
> >  }
> >  
> > -static bool _is_already_opened(const char *path, int as_idx)
> > +/* in *err returns -errno or index to _opened_fds[] */
> > +static bool _is_already_opened(const char *path, int as_idx, int *err)
> >  {
> >  	struct stat new;
> >  
> >  	assert(as_idx < ARRAY_SIZE(_opened_fds));
> > -	assert(as_idx <= _opened_fds_count);
> > +	assert(as_idx >= 0);
> >  
> >  	/*
> >  	 * we cannot even stat the device, so it's of no use - let's claim it's
> >  	 * already opened
> >  	 */
> > -	if (igt_debug_on(stat(path, &new) != 0))
> > +	if (igt_debug_on(stat(path, &new) != 0)) {
> > +		if (err)
> > +			*err = -errno;
> > +
> >  		return true;
> > +	}
> > +
> > +	if (err)
> > +		*err = 0;
> > +
> > +	for (int i = 0, end = min(as_idx, _opened_fds_count); i < end; ++i) {
> 
> AFAICU from our talk, by limiting the scope to min(as_idx, _opened_fds_count) 
> you tried to address a different issue -- the as_idx-th card potentially 
> having been already opened before and added to the _opened_fds table, then 
> closed.
> 

Old code here was:
> > -	for (int i = 0; i < as_idx; ++i) {

It can be changed to

	for (int i = 0; i < _opened_fds_count; ++i) {

with additional code here:

        if (i == as_idx) continue;

as we want to keep compatibility with previous behaviour which allows

fd = drm_open_driver(DRIVER_XE);
fd2 = drm_open_driver(DRIVER_XE);

Also note that old code was actually limiting checks here,
after opening 0...3rd card, call to open card(1) will
actually only check _opened_fds[0].

> > +		if (_opened_fds[i].fd == -1)
> > +			continue;
> >  
> > -	for (int i = 0; i < as_idx; ++i) {
> >  		/* did we cross filesystem boundary? */
> >  		assert(_opened_fds[i].stat.st_dev == new.st_dev);
> >  
> > -		if (_opened_fds[i].stat.st_ino == new.st_ino)
> > +		if (_opened_fds[i].stat.st_ino == new.st_ino) {
> > +			if (err)
> > +				*err = i;
> > +
> >  			return true;
> > +		}
> >  	}
> >  
> >  	return false;
> > @@ -359,23 +378,37 @@ static bool _is_already_opened(const char *path, int as_idx)
> >  static int __search_and_open(const char *base, int offset, unsigned int chipset, int as_idx)
> >  {
> >  	const char *forced;
> > +	int err;
> >  
> >  	forced = forced_driver();
> >  	if (forced)
> >  		igt_debug("Force option used: Using driver %s\n", forced);
> >  
> > -	for (int i = 0; i < 16; i++) {
> > +	for (int i = 0, idx = -1; i < 16 && idx < as_idx; i++) {
> >  		char name[80];
> >  		int fd;
> >  
> >  		sprintf(name, "%s%u", base, i + offset);
> >  
> > -		if (_is_already_opened(name, as_idx))
> > -			continue;
> > +		if (_is_already_opened(name, as_idx, &err)) {
> > +			if (err < 0)
> > +				continue;
> > +
> > +			if (idx + 1 < as_idx) {
> > +				++idx;
> > +				continue;
> > +			}
> > +		}
> >  
> >  		fd = __drm_open_device(name, chipset);
> > -		if (fd != -1)
> > +		if (fd == -1)
> > +			continue;
> > +
> > +		++idx;
> > +		if (idx == as_idx)
> >  			return fd;
> > +
> > +		close(fd);
> >  	}
> >  
> >  	return -1;
> > @@ -486,7 +519,7 @@ static bool __get_card_for_nth_filter(int idx, struct igt_device_card *card)
> >   *   * idx-th filter (starting with 0, filters are semicolon separated) is used
> >   *   * if there is no idx-th filter, goto 2
> >   *   * first device maching the filter is selected
> > - *   * if it's already opened (for indexes = 0..idx-1) we fail with -1
> > + *   * if it's already opened (for indexes != idx) we fail with -1
> >   *   * otherwise open the device and return the fd
> >   *
> >   * 2. compatibility mode - open the first DRM device we can find that is not
> > @@ -531,6 +564,7 @@ int __drm_open_driver_another(int idx, int chipset)
> >  	if (chipset != DRIVER_VGEM && igt_device_filter_count() > idx) {
> >  		struct igt_device_card card;
> >  		bool found;
> > +		int err;
> >  
> >  		found = __get_card_for_nth_filter(idx, &card);
> >  
> > @@ -539,14 +573,19 @@ int __drm_open_driver_another(int idx, int chipset)
> >  			found = __get_card_for_nth_filter(idx, &card);
> >  		}
> >  
> > -		if (!found || !strlen(card.card))
> > +		if (!found || !strlen(card.card)) {
> >  			igt_warn("No card matches the filter! [%s]\n",
> >  				 igt_device_filter_get(idx));
> > -		else if (_is_already_opened(card.card, idx))
> > -			igt_warn("card maching filter %d is already opened\n", idx);
> > -		else
> > +		} else if (_is_already_opened(card.card, idx, &err)) {
> > +			if (err < 0)
> > +				igt_warn("cannot stat card maching filter %d errno=%d card=%s\n", idx, err, card.card);
> > +			else if (idx != err)
> > +				igt_warn("card maching filter %d is already opened as %d, card=%s\n", idx, err, card.card);
> > +			else
> 
> Please note that if the aforementioned loop inside _is_already_opened() is 
> limited to min(as_idx, _opened_fds_count) then the above else case is never 
> true since err can never reach idx, then the below assignment is dead code.
> 

You are right, I would remove it. I was thinking about change
here to drop idx != err check, see below.

> The issue with before opened and closed cards should be addressed in a 
> different way and with a separate patch, I believe, maybe by removing 
> cards from _opened_fds on close.
> 
> Thanks,
> Janusz
> 

That was never done maybe because tests were run only once and
mostly on one-card system. Also there are cases when dev wants to call
close() directly, not with drm_close_driver(). After return from this
function all _fds data is stale, especially after driver reload but
this is a seperate problem.
Imho use case for this was to open all cards in one go and test them.

Seperate issue would be to decide it we allow use of repeated card=0
like:

sudo IGT_DEVICE=pci:vendor=Intel,card=0\;pci:vendor=Intel,card=0 \
    build/tests/gem_basic --r multigpu-create-close

as now this will fail in igt_warn("card matching ... is already opened as ...")

Regards,
Kamil

> > +				fd = __open_driver_exact(card.card, chipset);
> > +		} else {
> >  			fd = __open_driver_exact(card.card, chipset);
> > -
> > +		}
> >  	} else {
> >  		/* no filter for device idx, let's open whatever is available */
> >  		fd = __open_driver("/dev/dri/card", 0, chipset, idx);
> > 
> 
> 
> 
> 


More information about the igt-dev mailing list