| Message ID | 56DD9E7E.6000406@codeaurora.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Sinan, On 03/07/2016 04:30 PM, Sinan Kaya wrote: > On 3/6/2016 11:09 PM, Eric Auger wrote: >>> #define module_vfio_reset_handler(compat, acpihid, reset) \ >>>> MODULE_ALIAS("vfio-reset:" acpihid? acpihid: compat); \ >>>> >>>> This way, we'll create an alias with one of the provided strings. >> What if you want to use vfio platform driver for HiDMA in dt mode? the >> HiDma reset module advertises both acpihid and dt compat support but an >> alias module will be created only for acpihid. Then I think we will not >> be able to load the reset module in dt mode with existing code. >> > > Right, it won't work. Now that we know what MODULE_ALIAS does, there is no > harm in doing this. I think we should do this. > > #define module_vfio_reset_handler(compat, acpihid, reset) \ > MODULE_ALIAS("vfio-reset:" compat); \ > MODULE_ALIAS("vfio-reset:" acpihid); \ > > If we prefer ACPI over DT, there is no guarantee that somebody can boot DT > kernel with ACPI kernel compilation option enabled. > > If one of these are null, then the module alias will be "vfio-reset:" > > Of course to make things prettier, we could use "NOT SUPPORTED" as a string > instead of NULL. then, the module alias will be vfio-reset: NOT SUPPORTED". > > We could go one step further, and do. > > #define module_vfio_reset_handler(compat, acpihid, reset) \ > MODULE_ALIAS("vfio-reset: dt: " compat); \ > MODULE_ALIAS("vfio-reset: acpi: " acpihid); \ My gut feeling is we must not create a dummy alias when the mode (dt/acpi) is not supported/tested. It fills the modinfo section with spurious data, aliases are visible to modinfo, ... So I personally foresee 2 solutions, 1) we create a single alias using acpihid if supported or compat if not. Then even in dt mode we try to load this module through the acpihid alias. Looks weird but should work. 2) We simply move the module alias declaration out of this macro (to the reset module itself), define 2 aliases in case both dt and acpi are supported & tested. My personal preference is 2 I think. Best Regards Eric > > and change the code below for this too. > > >> This could work however with some rework in vfio_platform_common.c. In >> vfio_platform_get_reset we should try to load the module using the >> acpihid if the module load using compat alias fails. In the look-up >> table we can find the acpihid corresponding to the dt compat. > > I was planning to submit this for the next review. Still, I don't want to > assume that acpihid is the only working option. Somebody can boot DT kernel. > > diff --git a/drivers/vfio/platform/vfio_platform_common.c b/drivers/vfio/platform/vfio_platform_common.c > index 42d7545..ba585ad 100644 > --- a/drivers/vfio/platform/vfio_platform_common.c > +++ b/drivers/vfio/platform/vfio_platform_common.c > @@ -58,16 +58,30 @@ static vfio_platform_reset_fn_t vfio_platform_lookup_reset(const char *compat, > return reset_fn; > } > > -static void vfio_platform_get_reset(struct vfio_platform_device *vdev) > +static int vfio_platform_get_reset(struct vfio_platform_device *vdev) > { > + int rc; > + > vdev->reset = vfio_platform_lookup_reset(vdev->compat, vdev->acpihid, > &vdev->reset_module); > - if (!vdev->reset) { > - request_module("vfio-reset:%s", vdev->compat); > - vdev->reset = vfio_platform_lookup_reset(vdev->compat, > - vdev->acpihid, > - &vdev->reset_module); > - } > + if (vdev->reset) > + return 0; > + > + if (vdev->acpihid) > + rc = request_module("vfio-reset:%s", vdev->acpihid); > + > + if (rc && vdev->compat) > + rc = request_module("vfio-reset:%s", vdev->compat); > + > + if (rc) > + return rc; > + > + vdev->reset = vfio_platform_lookup_reset(vdev->compat, vdev->acpihid, > + &vdev->reset_module); > + if (vdev->reset) > + return 0; > + > + return -ENODEV; > } > > static void vfio_platform_put_reset(struct vfio_platform_device *vdev) > @@ -620,7 +634,11 @@ int vfio_platform_probe_common(struct vfio_platform_device *vdev, > return ret; > } > > - vfio_platform_get_reset(vdev); > + ret = vfio_platform_get_reset(vdev); > + if (ret) { > + iommu_group_put(group); > + return ret; > + } > > mutex_init(&vdev->igate); > > > Let me know what you think. > >
On 3/7/2016 11:46 PM, Eric Auger wrote: >> #define module_vfio_reset_handler(compat, acpihid, reset) \ >> > MODULE_ALIAS("vfio-reset: dt: " compat); \ >> > MODULE_ALIAS("vfio-reset: acpi: " acpihid); \ > My gut feeling is we must not create a dummy alias when the mode > (dt/acpi) is not supported/tested. It fills the modinfo section with > spurious data, aliases are visible to modinfo, ... > > So I personally foresee 2 solutions, > 1) we create a single alias using acpihid if supported or compat if not. > Then even in dt mode we try to load this module through the acpihid > alias. Looks weird but should work. > 2) We simply move the module alias declaration out of this macro (to the > reset module itself), define 2 aliases in case both dt and acpi are > supported & tested. > > My personal preference is 2 I think. > > Best Regards > > Eric > I agree with your assessment of #2. We are trying to be too smart. I'll post a patch with #2 solution.
Hi, On 3/8/2016 12:07 AM, Sinan Kaya wrote: > On 3/7/2016 11:46 PM, Eric Auger wrote: >>> #define module_vfio_reset_handler(compat, acpihid, reset) \ >>>> MODULE_ALIAS("vfio-reset: dt: " compat); \ >>>> MODULE_ALIAS("vfio-reset: acpi: " acpihid); \ >> My gut feeling is we must not create a dummy alias when the mode >> (dt/acpi) is not supported/tested. It fills the modinfo section with >> spurious data, aliases are visible to modinfo, ... >> >> So I personally foresee 2 solutions, >> 1) we create a single alias using acpihid if supported or compat if not. >> Then even in dt mode we try to load this module through the acpihid >> alias. Looks weird but should work. >> 2) We simply move the module alias declaration out of this macro (to the >> reset module itself), define 2 aliases in case both dt and acpi are >> supported & tested. >> >> My personal preference is 2 I think. >> >> Best Regards >> >> Eric >> > > I agree with your assessment of #2. We are trying to be too smart. > I'll post a patch with #2 solution. > I split patch 8 and 9 from the series and posted a new set here. http://www.spinics.net/lists/arm-kernel/msg489578.html
diff --git a/drivers/vfio/platform/vfio_platform_common.c b/drivers/vfio/platform/vfio_platform_common.c index 42d7545..ba585ad 100644 --- a/drivers/vfio/platform/vfio_platform_common.c +++ b/drivers/vfio/platform/vfio_platform_common.c @@ -58,16 +58,30 @@ static vfio_platform_reset_fn_t vfio_platform_lookup_reset(const char *compat, return reset_fn; } -static void vfio_platform_get_reset(struct vfio_platform_device *vdev) +static int vfio_platform_get_reset(struct vfio_platform_device *vdev) { + int rc; + vdev->reset = vfio_platform_lookup_reset(vdev->compat, vdev->acpihid, &vdev->reset_module); - if (!vdev->reset) { - request_module("vfio-reset:%s", vdev->compat); - vdev->reset = vfio_platform_lookup_reset(vdev->compat, - vdev->acpihid, - &vdev->reset_module); - } + if (vdev->reset) + return 0; + + if (vdev->acpihid) + rc = request_module("vfio-reset:%s", vdev->acpihid); + + if (rc && vdev->compat) + rc = request_module("vfio-reset:%s", vdev->compat); + + if (rc) + return rc; + + vdev->reset = vfio_platform_lookup_reset(vdev->compat, vdev->acpihid, + &vdev->reset_module); + if (vdev->reset) + return 0; + + return -ENODEV; } static void vfio_platform_put_reset(struct vfio_platform_device *vdev) @@ -620,7 +634,11 @@ int vfio_platform_probe_common(struct vfio_platform_device *vdev, return ret; } - vfio_platform_get_reset(vdev); + ret = vfio_platform_get_reset(vdev); + if (ret) { + iommu_group_put(group); + return ret; + } mutex_init(&vdev->igate);