| Message ID | 20240409233942.828440-1-peter.colberg@intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | fpga: dfl: fix kernel warning on port release/assign for SRIOV | expand |
On Tue, Apr 09, 2024 at 07:39:33PM -0400, Peter Colberg wrote: > DFL ports are registered as platform devices in PF mode. The port device > should be removed from the host when the user wants to configure the > port as a VF and pass through to a virtual machine. The FME device > ioctls DFL_FPGA_FME_PORT_RELEASE/ASSIGN are designed for this purpose. > > In the previous implementation, the port platform device is not completely > destroyed on port release: it is removed from the system by > platform_device_del(), but the platform device instance is retained. > When the port assign ioctl is called, the platform device is added back by > platform_device_add(), which conflicts with this comment of device_add(): > "Do not call this routine more than once for any device structure", and > will cause a kernel warning at runtime. > > This patch tries to completely unregister the port platform device on > release and registers a new one on assign. But the main work is to remove > the dependency on struct dfl_feature_platform_data for many internal DFL > APIs. This structure holds many DFL enumeration infos for feature devices. > Many DFL APIs are expected to work with these info even when the port > platform device is unregistered. But with the change the platform_data will > be freed in this case. So this patch introduces a new structure > dfl_feature_dev_data for these APIs, which acts similarly to the previous > dfl_feature_platform_data. The dfl_feature_platform_data then only needs a > pointer to dfl_feature_dev_data to make the feature device driver work. > > The single monolithic v1 patch is split into multiple, smaller patches > at the request of the maintainer. The first patch adds temporary macros > that alias dfl_feature_dev_data ("fdata") to dfl_feature_platform_data > ("pdata") and associated functions from the "fdata" to the corresponding > "pdata" variants. Subsequent patches separate out most of the symbol > name changes required by this patch series, one patch per file. The last One patch per file is not a requirement, simple replacement across multiple files won't cause trouble for reviewers. The important thing is that don't bury the key changes in these symbol replacement so that people can't get the point.
On Tue, 2024-04-23 at 16:27 +0800, Xu Yilun wrote: > On Tue, Apr 09, 2024 at 07:39:33PM -0400, Peter Colberg wrote: > > DFL ports are registered as platform devices in PF mode. The port device > > should be removed from the host when the user wants to configure the > > port as a VF and pass through to a virtual machine. The FME device > > ioctls DFL_FPGA_FME_PORT_RELEASE/ASSIGN are designed for this purpose. > > > > In the previous implementation, the port platform device is not completely > > destroyed on port release: it is removed from the system by > > platform_device_del(), but the platform device instance is retained. > > When the port assign ioctl is called, the platform device is added back by > > platform_device_add(), which conflicts with this comment of device_add(): > > "Do not call this routine more than once for any device structure", and > > will cause a kernel warning at runtime. > > > > This patch tries to completely unregister the port platform device on > > release and registers a new one on assign. But the main work is to remove > > the dependency on struct dfl_feature_platform_data for many internal DFL > > APIs. This structure holds many DFL enumeration infos for feature devices. > > Many DFL APIs are expected to work with these info even when the port > > platform device is unregistered. But with the change the platform_data will > > be freed in this case. So this patch introduces a new structure > > dfl_feature_dev_data for these APIs, which acts similarly to the previous > > dfl_feature_platform_data. The dfl_feature_platform_data then only needs a > > pointer to dfl_feature_dev_data to make the feature device driver work. > > > > The single monolithic v1 patch is split into multiple, smaller patches > > at the request of the maintainer. The first patch adds temporary macros > > that alias dfl_feature_dev_data ("fdata") to dfl_feature_platform_data > > ("pdata") and associated functions from the "fdata" to the corresponding > > "pdata" variants. Subsequent patches separate out most of the symbol > > name changes required by this patch series, one patch per file. The last > > One patch per file is not a requirement, simple replacement across > multiple files won't cause trouble for reviewers. The important thing is > that don't bury the key changes in these symbol replacement so that > people can't get the point. Thank you, in hindsight this should have been obvious. The v3 patch series now breaks the v1 patch into logical, self-contained patches. Thanks, Peter
DFL ports are registered as platform devices in PF mode. The port device should be removed from the host when the user wants to configure the port as a VF and pass through to a virtual machine. The FME device ioctls DFL_FPGA_FME_PORT_RELEASE/ASSIGN are designed for this purpose. In the previous implementation, the port platform device is not completely destroyed on port release: it is removed from the system by platform_device_del(), but the platform device instance is retained. When the port assign ioctl is called, the platform device is added back by platform_device_add(), which conflicts with this comment of device_add(): "Do not call this routine more than once for any device structure", and will cause a kernel warning at runtime. This patch tries to completely unregister the port platform device on release and registers a new one on assign. But the main work is to remove the dependency on struct dfl_feature_platform_data for many internal DFL APIs. This structure holds many DFL enumeration infos for feature devices. Many DFL APIs are expected to work with these info even when the port platform device is unregistered. But with the change the platform_data will be freed in this case. So this patch introduces a new structure dfl_feature_dev_data for these APIs, which acts similarly to the previous dfl_feature_platform_data. The dfl_feature_platform_data then only needs a pointer to dfl_feature_dev_data to make the feature device driver work. The single monolithic v1 patch is split into multiple, smaller patches at the request of the maintainer. The first patch adds temporary macros that alias dfl_feature_dev_data ("fdata") to dfl_feature_platform_data ("pdata") and associated functions from the "fdata" to the corresponding "pdata" variants. Subsequent patches separate out most of the symbol name changes required by this patch series, one patch per file. The last patch of the series removes the macros and applies the actual change. Link: https://lore.kernel.org/all/DM6PR11MB3819F9CCD0A6126B55BCB47685FB9@DM6PR11MB3819.namprd11.prod.outlook.com/T/#t This series applies after the following non-RFC patches: - fpga: dfl: remove unused function is_dfl_feature_present - [v2] fpga: dfl: remove unused member pdata from struct dfl_{afu,fme} Changes since v1: - Split monolithic patch into series at request of maintainer. - Substitute binfo->type for removed function feature_dev_id_type() in parse_feature_irqs(). - Return ERR_PTR(-ENOMEM) on !feature->params in binfo_create_feature_dev_data(). - Reorder cdev as first member of struct dfl_feature_platform_data such that container_of() to obtain pdata evaluates to a no-op. - Change afu_ioctl_*() to receive dfl_feature_dev_data instead of dfl_feature_platform_data. - Change fme_hdr_ioctl_*() to receive dfl_feature_dev_data instead of dfl_feature_platform_data. - Replace local variable pdata with fdata in afu_mmap(). - Remove unused local variable pdata in afu_dev_{init,destroy}(). - Remove unused local variable pdata in fme_dev_{init,destroy}(). - Reorder local variables in afu_dma_unpin_pages() to reverse Christmas tree order. - Align kernel-doc function name for __dfl_fpga_cdev_find_port_data(). - Substitute @fdata for @pdata in kernel-doc comments for dfl_fme_create_mgr() and dfl_fme_destroy_mgr(). Peter Colberg (8): fpga: dfl: alias dfl_feature_dev_data to dfl_feature_platform_data fpga: dfl: migrate AFU DMA region management driver to dfl_feature_dev_data fpga: dfl: migrate AFU MMIO region management driver to dfl_feature_dev_data fpga: dfl: migrate FPGA Management Engine driver to dfl_feature_dev_data fpga: dfl: migrate FME partial reconfiguration driver to dfl_feature_dev_data fpga: dfl: migrate Accelerated Function Unit driver to dfl_feature_dev_data fpga: dfl: migrate DFL support header to dfl_feature_dev_data fpga: dfl: migrate dfl_get_feature_by_id() to dfl_feature_dev_data Xu Yilun (1): fpga: dfl: fix kernel warning on port release/assign for SRIOV drivers/fpga/dfl-afu-dma-region.c | 119 ++++---- drivers/fpga/dfl-afu-error.c | 59 ++-- drivers/fpga/dfl-afu-main.c | 281 +++++++++---------- drivers/fpga/dfl-afu-region.c | 51 ++-- drivers/fpga/dfl-afu.h | 26 +- drivers/fpga/dfl-fme-br.c | 24 +- drivers/fpga/dfl-fme-error.c | 98 +++---- drivers/fpga/dfl-fme-main.c | 86 +++--- drivers/fpga/dfl-fme-pr.c | 88 +++--- drivers/fpga/dfl.c | 433 +++++++++++++++--------------- drivers/fpga/dfl.h | 139 ++++++---- 11 files changed, 722 insertions(+), 682 deletions(-)