[Intel-gfx] [PATCH v3 i-g-t] lib/intel_mmio: Fix mmapped resources not unmapped on fini
Janusz Krzysztofik
janusz.krzysztofik at linux.intel.com
Mon Mar 7 15:06:10 UTC 2022
Hi Kamil,
On Monday, 7 March 2022 14:23:30 CET Kamil Konieczny wrote:
> Hi Janusz,
>
> Dnia 2022-03-07 at 09:26:43 +0100, Janusz Krzysztofik napisał(a):
> > Commit 5f3cfa485eb4 ("lib: Use safe wrappers around libpciaccess
> > initialization functions") took care of not leaking memory allocated by
> > pci_system_init() but didn't take care of users potentially attempting to
> > reinitialize global data maintained by libpciaccess. For example,
> > intel_register_access_init() mmaps device's PCI BAR0 resource with
> > pci_device_map_range() but intel_register_access_fini() doesn't unmap it
> > and next call to intel_register_access_init() fails on attempt to mmap it
> > again.
> >
> > Fix it, and also provide intel_mmio_unmap_*() counterparts to public
> > functions intel_mmio_use_pci_bar() and intel_mmio_use_dump_file().
> >
> > v2: apply last minute fixes, cached but unfortunately not committed before
> > sending
> > v3: use .pci_device_id field content as an indicator of arg initialization
> > via intel_register_access_init(),
> > - improve checks of argument initialization status,
> > - shorten warning messages (Kamil),
> > - don't fill .mmio_size field until initialization succeeds (Kamil)
> >
> > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > Cc: Kamil Konieczny <kamil.konieczny at linux.intel.com>
> > ---
> > lib/intel_io.h | 4 +++
> > lib/intel_mmio.c | 64 +++++++++++++++++++++++++++++++++++++++++++++---
> > 2 files changed, 65 insertions(+), 3 deletions(-)
> >
> > diff --git a/lib/intel_io.h b/lib/intel_io.h
> > index 1cfe4fb6b9..ea2649d9bc 100644
> > --- a/lib/intel_io.h
> > +++ b/lib/intel_io.h
> > @@ -49,6 +49,8 @@ struct intel_register_map {
> >
> > struct intel_mmio_data {
> > void *igt_mmio;
> > + size_t mmio_size;
> > + struct pci_device *dev;
> > struct intel_register_map map;
> > uint32_t pci_device_id;
> > int key;
> > @@ -57,7 +59,9 @@ struct intel_mmio_data {
> >
> > void intel_mmio_use_pci_bar(struct intel_mmio_data *mmio_data,
> > struct pci_device *pci_dev);
> > +void intel_mmio_unmap_pci_bar(struct intel_mmio_data *mmio_data);
> > void intel_mmio_use_dump_file(struct intel_mmio_data *mmio_data, char *file);
> > +void intel_mmio_unmap_dump_file(struct intel_mmio_data *mmio_data);
> >
> > int intel_register_access_init(struct intel_mmio_data *mmio_data,
> > struct pci_device *pci_dev, int safe, int fd);
> > diff --git a/lib/intel_mmio.c b/lib/intel_mmio.c
> > index 667a69f5aa..d6ce0ee3ea 100644
> > --- a/lib/intel_mmio.c
> > +++ b/lib/intel_mmio.c
> > @@ -82,6 +82,8 @@ void *igt_global_mmio;
> > * Sets also up mmio_data->igt_mmio to point at the data contained
> > * in @file. This allows the same code to get reused for dumping and decoding
> > * from running hardware as from register dumps.
> > + *
> > + * Users are expected to call intel_mmio_unmap_dump_file() after use.
> > */
> > void
> > intel_mmio_use_dump_file(struct intel_mmio_data *mmio_data, char *file)
> > @@ -99,11 +101,32 @@ intel_mmio_use_dump_file(struct intel_mmio_data *mmio_data, char *file)
> > igt_fail_on_f(mmio_data->igt_mmio == MAP_FAILED,
> > "Couldn't mmap %s\n", file);
> >
> > + mmio_data->mmio_size = st.st_size;
> > igt_global_mmio = mmio_data->igt_mmio;
> >
> > close(fd);
> > }
> >
> > +/**
> > + * intel_mmio_unmap_dump_file:
> > + * @mmio_data: mmio structure for IO operations
> > + *
> > + * Unmaps a dump file mmapped with intel_mmio_use_dump_file()
> > + */
> > +void intel_mmio_unmap_dump_file(struct intel_mmio_data *mmio_data)
> > +{
> > + if (igt_warn_on_f(mmio_data->dev,
> > + "test bug: arg initialized with intel_mmio_use_pci_bar()\n"))
> > + return;
>
> Please add a global description about this kind of errors, this
> one is for using unmap when mmio was mmap-ed from other mmap
> type.
Can you please be more specific in what you mean by "global description of
this kind of errors"? A more detailed warning? A comment? If the latter
then how would you like me to make it global?
If you just don't like the reference to intel_mmio_use_pci_bar() here then
would you be satisfied with something like "test bug: arg initialized by a
method other than intel_mmio_use_dump_file()\n"?
> > + if (igt_warn_on_f(!mmio_data->mmio_size,
> > + "test bug: arg not initialized\n"))
> > + return;
>
> Can we replace this with one check igt_global_mmio != NULL ?
> Something like:
>
> if (igt_warn_on_f(!igt_global_mmio,
> "mmio regs not mmap-ed\n"))
> return;
>
> Or should we add this before all other checks in unmap functions
> and keep this additional check ?
Why igt_global_mmio again? I still think this global variable is broken and
users should just use the structure they pass to intel_mmio_use_*() or
intel_register_access_init(), then introducing another a dependency on it with
this patch doesn't make sense to me. If you think the opposite then please
explain why.
> > +
> > + igt_global_mmio = NULL;
> > + igt_debug_on(munmap(mmio_data->igt_mmio, mmio_data->mmio_size) < 0);
> > + mmio_data->mmio_size = 0;
> > +}
> > +
> > /**
> > * intel_mmio_use_pci_bar:
> > * @mmio_data: mmio structure for IO operations
> > @@ -112,6 +135,8 @@ intel_mmio_use_dump_file(struct intel_mmio_data *mmio_data, char *file)
> > * Fill a mmio_data stucture with igt_mmio to point at the mmio bar.
> > *
> > * @pci_dev can be obtained from intel_get_pci_device().
> > + *
> > + * Users are expected to call intel_mmio_unmap_pci_bar() after use.
> > */
> > void
> > intel_mmio_use_pci_bar(struct intel_mmio_data *mmio_data, struct pci_device *pci_dev)
> > @@ -141,10 +166,34 @@ intel_mmio_use_pci_bar(struct intel_mmio_data *mmio_data, struct pci_device *pci
> > PCI_DEV_MAP_FLAG_WRITABLE,
> > &mmio_data->igt_mmio);
> >
> > - igt_global_mmio = mmio_data->igt_mmio;
> > -
> > igt_fail_on_f(error != 0,
> > "Couldn't map MMIO region\n");
> > +
> > + mmio_data->mmio_size = mmio_size;
> > + mmio_data->dev = pci_dev;
> > + igt_global_mmio = mmio_data->igt_mmio;
To be consequent with what I said above, in next version of the patch I'm
going to not touch the assignment of mmio_data->igt_mmio to the out of scope
and depreciated igt_global_mmio, leaving it as broken as it already is.
> > +}
> > +
> > +/**
> > + * intel_mmio_unmap_pci_bar:
> > + * @mmio_data: mmio structure for IO operations
> > + *
> > + * Unmaps a PCI BAR region mmapped with intel_mmio_use_pci_bar()
> > + */
> > +void intel_mmio_unmap_pci_bar(struct intel_mmio_data *mmio_data)
> > +{
> > + if (igt_warn_on_f(mmio_data->pci_device_id,
> > + "test bug: arg initialized with intel_register_access_init()\n"))
Similarly to the case of intel_mmio_unmap_dump_file(), I can change the
message to "test bug: arg initialized with a method other than
intel_mmio_use_pci_bar()\n" if that's what you prefer.
> > + return;
> > + if (igt_warn_on_f(!mmio_data->dev,
> > + "test bug: arg not initialized with intel_mmio_use_pci_bar()\n"))
> > + return;
> > +
> > + igt_global_mmio = NULL;
> > + igt_debug_on(pci_device_unmap_range(mmio_data->dev,
> > + mmio_data->igt_mmio, mmio_data->mmio_size) < 0);
> > + mmio_data->dev = NULL;
> > + mmio_data->mmio_size = 0;
> > }
> >
> > static void
> > @@ -166,6 +215,8 @@ release_forcewake_lock(int fd)
> > * It also initializes mmio_data->igt_mmio like intel_mmio_use_pci_bar().
> > *
> > * @pci_dev can be obtained from intel_get_pci_device().
> > + *
> > + * Users are expected to call intel_register_access_fini() after use.
> > */
> > int
> > intel_register_access_init(struct intel_mmio_data *mmio_data, struct pci_device *pci_dev, int safe, int fd)
> > @@ -222,8 +273,15 @@ int intel_register_access_needs_fakewake(struct intel_mmio_data *mmio_data)
> > void
> > intel_register_access_fini(struct intel_mmio_data *mmio_data)
> > {
> > - if (mmio_data->key && intel_register_access_needs_wake(mmio_data))
> > + if (igt_warn_on_f(!mmio_data->pci_device_id,
> > + "test bug: arg not initialized with intel_register_access_init()\n"))
> > + return;
> > +
> > + if (intel_register_access_needs_wake(mmio_data))
As mmio_data->key is no longer used since v3 as an indication of mmio_data
having been initialized by intel_register_access_init() or not, the original
(potentially broken) condition:
if (mmio_data->key && intel_register_access_needs_wake(mmio_data))
should be not touched by this patch as out of scope but its original form
preserved. I'm going to restore it in next version of the patch.
> > release_forcewake_lock(mmio_data->key);
> > +
> > + mmio_data->pci_device_id = 0;
>
> Here we should check other conditions so no warn triggers in unmap_pci_bar
> or make the messages generic (and document it in comments at beggining)
> or maybe make helper with no checks for unmap_pci_bar.
Why? I can't see any potential scenario where mmio_data->pci_device_id is not
0 but mmio_data->dev is NULL. If you can see one then please tell me more
about it.
Thanks,
Janusz
> > + intel_mmio_unmap_pci_bar(mmio_data);
> > }
> >
> > /**
>
More information about the Intel-gfx
mailing list