| Message ID | 20170801114926.1171418-1-arnd@arndb.de |
|---|---|
| State | New |
| Headers | show |
On Tue, Aug 01, 2017 at 01:48:48PM +0200, Arnd Bergmann wrote: > Removing the btt_rw_page/pmem_rw_page functions had a surprising > side-effect of introducing a false-positive warning in another > function, due to changed inlining decisions in gcc: > > In file included from drivers/nvdimm/pmem.c:36:0: > drivers/nvdimm/pmem.c: In function 'pmem_make_request': > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > drivers/nvdimm/pmem.c:174:16: note: 'start' was declared here > In file included from drivers/nvdimm/btt.c:27:0: > drivers/nvdimm/btt.c: In function 'btt_make_request': > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > drivers/nvdimm/btt.c:1202:16: note: 'start' was declared here > > The problem is that gcc fails to track the value of the 'do_acct' > variable here and has to read it back from stack, but it does > remember that 'start' may be uninitialized sometimes. > > This shuts up the warning by making nd_iostat_start() always > initialize the 'start' variable. In those cases that gcc successfully > tracks the state of the variable, this will have no effect. > > Fixes: 503a5e89b1de ("drivers/nvdimm/btt.c: remove btt_rw_page()") > Fixes: 58100d6e735e ("drivers/nvdimm/pmem.c: remove pmem_rw_page()") > Signed-off-by: Arnd Bergmann <arnd@arndb.de> This change looks fine: Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com> I believe the patches removing the btt_rw_page() and btt_rw_page() are on hold until I can get some performance numbers to justify them. Dan, do you want to take this as is, or do you want me to include it in my larger rw_page() series if/when that gets revived?
On Tue, Aug 1, 2017 at 11:00 AM, Ross Zwisler <ross.zwisler@linux.intel.com> wrote: > On Tue, Aug 01, 2017 at 01:48:48PM +0200, Arnd Bergmann wrote: >> Removing the btt_rw_page/pmem_rw_page functions had a surprising >> side-effect of introducing a false-positive warning in another >> function, due to changed inlining decisions in gcc: >> >> In file included from drivers/nvdimm/pmem.c:36:0: >> drivers/nvdimm/pmem.c: In function 'pmem_make_request': >> drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] >> drivers/nvdimm/pmem.c:174:16: note: 'start' was declared here >> In file included from drivers/nvdimm/btt.c:27:0: >> drivers/nvdimm/btt.c: In function 'btt_make_request': >> drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] >> drivers/nvdimm/btt.c:1202:16: note: 'start' was declared here >> >> The problem is that gcc fails to track the value of the 'do_acct' >> variable here and has to read it back from stack, but it does >> remember that 'start' may be uninitialized sometimes. >> >> This shuts up the warning by making nd_iostat_start() always >> initialize the 'start' variable. In those cases that gcc successfully >> tracks the state of the variable, this will have no effect. >> >> Fixes: 503a5e89b1de ("drivers/nvdimm/btt.c: remove btt_rw_page()") >> Fixes: 58100d6e735e ("drivers/nvdimm/pmem.c: remove pmem_rw_page()") >> Signed-off-by: Arnd Bergmann <arnd@arndb.de> > > This change looks fine: > > Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com> > > I believe the patches removing the btt_rw_page() and btt_rw_page() are on hold > until I can get some performance numbers to justify them. > > Dan, do you want to take this as is, or do you want me to include it in my > larger rw_page() series if/when that gets revived? I'd say include it with your set.
On Tue, 1 Aug 2017 13:48:48 +0200 Arnd Bergmann <arnd@arndb.de> wrote: > Removing the btt_rw_page/pmem_rw_page functions had a surprising > side-effect of introducing a false-positive warning in another > function, due to changed inlining decisions in gcc: > > In file included from drivers/nvdimm/pmem.c:36:0: > drivers/nvdimm/pmem.c: In function 'pmem_make_request': > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > drivers/nvdimm/pmem.c:174:16: note: 'start' was declared here > In file included from drivers/nvdimm/btt.c:27:0: > drivers/nvdimm/btt.c: In function 'btt_make_request': > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > drivers/nvdimm/btt.c:1202:16: note: 'start' was declared here > > The problem is that gcc fails to track the value of the 'do_acct' > variable here and has to read it back from stack, but it does > remember that 'start' may be uninitialized sometimes. > > This shuts up the warning by making nd_iostat_start() always > initialize the 'start' variable. In those cases that gcc successfully > tracks the state of the variable, this will have no effect. > > ... > > --- a/drivers/nvdimm/nd.h > +++ b/drivers/nvdimm/nd.h > @@ -392,8 +392,10 @@ static inline bool nd_iostat_start(struct bio *bio, unsigned long *start) > { > struct gendisk *disk = bio->bi_bdev->bd_disk; > > - if (!blk_queue_io_stat(disk->queue)) > + if (!blk_queue_io_stat(disk->queue)) { > + *start = 0; > return false; > + } > > *start = jiffies; > generic_start_io_acct(bio_data_dir(bio), Well that's sad. The future of btt-remove-btt_rw_page.patch and friends is shrouded in mystery, but if we proceed that way then yes, I guess we'll need to work around such gcc glitches. But let's not leave apparently-unneeded code in place without telling people why it is in fact needed?--- a/drivers/nvdimm/nd.h~nvdimm-avoid-bogus-wmaybe-uninitialized-warning-fix +++ a/drivers/nvdimm/nd.h @@ -393,7 +393,7 @@ static inline bool nd_iostat_start(struc struct gendisk *disk = bio->bi_bdev->bd_disk; if (!blk_queue_io_stat(disk->queue)) { - *start = 0; + *start = 0; /* Suppress bogus warning */ return false; }
On Tue, Aug 01, 2017 at 02:45:34PM -0700, Andrew Morton wrote: > On Tue, 1 Aug 2017 13:48:48 +0200 Arnd Bergmann <arnd@arndb.de> wrote: > > > Removing the btt_rw_page/pmem_rw_page functions had a surprising > > side-effect of introducing a false-positive warning in another > > function, due to changed inlining decisions in gcc: > > > > In file included from drivers/nvdimm/pmem.c:36:0: > > drivers/nvdimm/pmem.c: In function 'pmem_make_request': > > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > > drivers/nvdimm/pmem.c:174:16: note: 'start' was declared here > > In file included from drivers/nvdimm/btt.c:27:0: > > drivers/nvdimm/btt.c: In function 'btt_make_request': > > drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] > > drivers/nvdimm/btt.c:1202:16: note: 'start' was declared here > > > > The problem is that gcc fails to track the value of the 'do_acct' > > variable here and has to read it back from stack, but it does > > remember that 'start' may be uninitialized sometimes. > > > > This shuts up the warning by making nd_iostat_start() always > > initialize the 'start' variable. In those cases that gcc successfully > > tracks the state of the variable, this will have no effect. > > > > ... > > > > --- a/drivers/nvdimm/nd.h > > +++ b/drivers/nvdimm/nd.h > > @@ -392,8 +392,10 @@ static inline bool nd_iostat_start(struct bio *bio, unsigned long *start) > > { > > struct gendisk *disk = bio->bi_bdev->bd_disk; > > > > - if (!blk_queue_io_stat(disk->queue)) > > + if (!blk_queue_io_stat(disk->queue)) { > > + *start = 0; > > return false; > > + } > > > > *start = jiffies; > > generic_start_io_acct(bio_data_dir(bio), > > Well that's sad. > > The future of btt-remove-btt_rw_page.patch and friends is shrouded in > mystery, but if we proceed that way then yes, I guess we'll need to > work around such gcc glitches. > > But let's not leave apparently-unneeded code in place without telling > people why it is in fact needed? Maybe it's just cleaner to initialize 'start' in all the callers, so we don't have a mysterious line and have to remember why it's there / comment it? I'll throw a patch like that in my series if/when I repost.
On Wed, Aug 2, 2017 at 12:23 AM, Ross Zwisler <ross.zwisler@linux.intel.com> wrote: > On Tue, Aug 01, 2017 at 02:45:34PM -0700, Andrew Morton wrote: >> On Tue, 1 Aug 2017 13:48:48 +0200 Arnd Bergmann <arnd@arndb.de> wrote: >> > --- a/drivers/nvdimm/nd.h >> > +++ b/drivers/nvdimm/nd.h >> > @@ -392,8 +392,10 @@ static inline bool nd_iostat_start(struct bio *bio, unsigned long *start) >> > { >> > struct gendisk *disk = bio->bi_bdev->bd_disk; >> > >> > - if (!blk_queue_io_stat(disk->queue)) >> > + if (!blk_queue_io_stat(disk->queue)) { >> > + *start = 0; >> > return false; >> > + } >> > >> > *start = jiffies; >> > generic_start_io_acct(bio_data_dir(bio), >> >> Well that's sad. >> >> The future of btt-remove-btt_rw_page.patch and friends is shrouded in >> mystery, but if we proceed that way then yes, I guess we'll need to >> work around such gcc glitches. >> >> But let's not leave apparently-unneeded code in place without telling >> people why it is in fact needed? > > Maybe it's just cleaner to initialize 'start' in all the callers, so we don't > have a mysterious line and have to remember why it's there / comment it? I considered that but decided that would be worse, since it shuts up more potential warnings about actual uninitialized use of the variable, and is slightly harder for the compiler to optimize away. You also end up having to add a comment in multiple places. Note that Andrew already added a comment when he applied my patch to his mmotm tree. Arnd
diff --git a/drivers/nvdimm/nd.h b/drivers/nvdimm/nd.h index e1b5715bd91f..64f79a156456 100644 --- a/drivers/nvdimm/nd.h +++ b/drivers/nvdimm/nd.h @@ -392,8 +392,10 @@ static inline bool nd_iostat_start(struct bio *bio, unsigned long *start) { struct gendisk *disk = bio->bi_bdev->bd_disk; - if (!blk_queue_io_stat(disk->queue)) + if (!blk_queue_io_stat(disk->queue)) { + *start = 0; return false; + } *start = jiffies; generic_start_io_acct(bio_data_dir(bio),
Removing the btt_rw_page/pmem_rw_page functions had a surprising side-effect of introducing a false-positive warning in another function, due to changed inlining decisions in gcc: In file included from drivers/nvdimm/pmem.c:36:0: drivers/nvdimm/pmem.c: In function 'pmem_make_request': drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] drivers/nvdimm/pmem.c:174:16: note: 'start' was declared here In file included from drivers/nvdimm/btt.c:27:0: drivers/nvdimm/btt.c: In function 'btt_make_request': drivers/nvdimm/nd.h:407:2: error: 'start' may be used uninitialized in this function [-Werror=maybe-uninitialized] drivers/nvdimm/btt.c:1202:16: note: 'start' was declared here The problem is that gcc fails to track the value of the 'do_acct' variable here and has to read it back from stack, but it does remember that 'start' may be uninitialized sometimes. This shuts up the warning by making nd_iostat_start() always initialize the 'start' variable. In those cases that gcc successfully tracks the state of the variable, this will have no effect. Fixes: 503a5e89b1de ("drivers/nvdimm/btt.c: remove btt_rw_page()") Fixes: 58100d6e735e ("drivers/nvdimm/pmem.c: remove pmem_rw_page()") Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- drivers/nvdimm/nd.h | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) -- 2.9.0