| Message ID | 20220630051630.1718927-1-anshuman.khandual@arm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm/mmap: Drop __SXXX/__PXXX macros from across platforms | expand |
On 6/30/22 10:46, Anshuman Khandual wrote: > __SXXX/__PXXX macros is an unnecessary abstraction layer in creating the > generic protection_map[] array which is used for vm_get_page_prot(). This > abstraction layer can be avoided, if the platforms just define the array > protection_map[] for all possible vm_flags access permission combinations > and also export vm_get_page_prot() implementation. > > This series drops __SXXX/__PXXX macros from across platforms in the tree. > First it build protects generic protection_map[] array with '#ifdef __P000' > and moves it inside platforms which enable ARCH_HAS_VM_GET_PAGE_PROT. Later > this build protects same array with '#ifdef ARCH_HAS_VM_GET_PAGE_PROT' and > moves inside remaining platforms while enabling ARCH_HAS_VM_GET_PAGE_PROT. > This adds a new macro DECLARE_VM_GET_PAGE_PROT defining the current generic > vm_get_page_prot(), in order for it to be reused on platforms that do not > require custom implementation. Finally, ARCH_HAS_VM_GET_PAGE_PROT can just > be dropped, as all platforms now define and export vm_get_page_prot(), via > looking up a private and static protection_map[] array. protection_map[] > data type has been changed as 'static const' on all platforms that do not > change it during boot. > > This series applies on v5.19-rc4 and has been build tested for multiple > platforms. While here it has dropped off all previous tags from folks after > the current restructuring. Series common CC list has been expanded to cover > all impacted platforms for wider reach. > > - Anshuman > > Changes in V6: > > - Converted protection_map[] array as 'static const' on sparc32 platform > - Rebased on v5.19-rc4 > - Collected tags There are two linux-next based build fixes for this series (listed below), when vm_get_page_prot() gets redefined with !CONFIG_MMU. Platform vm_get_page_prot() is required only with CONFIG_MMU enabled, otherwise there is a generic fallback stub in include/linux/mm.h https://lore.kernel.org/all/20220705221411.3381797-1-jcmvbkbc@gmail.com/ [xtensa] https://lore.kernel.org/all/20220706054002.1936820-1-anshuman.khandual@arm.com/ [sh] It does not seem CONFIG_MMU can be disabled on other platforms thus exposing a build failure. But just to be on the safer side, should all vm_get_page_prot() be wrapped around with #ifdef CONFIG_MMU ? In that case will resend the series with above build fixes folded back in as well. Please do suggest. Thank you. - Anshuman
Le 06/07/2022 à 07:57, Anshuman Khandual a écrit : > > > On 6/30/22 10:46, Anshuman Khandual wrote: >> __SXXX/__PXXX macros is an unnecessary abstraction layer in creating the >> generic protection_map[] array which is used for vm_get_page_prot(). This >> abstraction layer can be avoided, if the platforms just define the array >> protection_map[] for all possible vm_flags access permission combinations >> and also export vm_get_page_prot() implementation. >> >> This series drops __SXXX/__PXXX macros from across platforms in the tree. >> First it build protects generic protection_map[] array with '#ifdef __P000' >> and moves it inside platforms which enable ARCH_HAS_VM_GET_PAGE_PROT. Later >> this build protects same array with '#ifdef ARCH_HAS_VM_GET_PAGE_PROT' and >> moves inside remaining platforms while enabling ARCH_HAS_VM_GET_PAGE_PROT. >> This adds a new macro DECLARE_VM_GET_PAGE_PROT defining the current generic >> vm_get_page_prot(), in order for it to be reused on platforms that do not >> require custom implementation. Finally, ARCH_HAS_VM_GET_PAGE_PROT can just >> be dropped, as all platforms now define and export vm_get_page_prot(), via >> looking up a private and static protection_map[] array. protection_map[] >> data type has been changed as 'static const' on all platforms that do not >> change it during boot. >> >> This series applies on v5.19-rc4 and has been build tested for multiple >> platforms. While here it has dropped off all previous tags from folks after >> the current restructuring. Series common CC list has been expanded to cover >> all impacted platforms for wider reach. >> >> - Anshuman >> >> Changes in V6: >> >> - Converted protection_map[] array as 'static const' on sparc32 platform >> - Rebased on v5.19-rc4 >> - Collected tags > > There are two linux-next based build fixes for this series (listed below), when > vm_get_page_prot() gets redefined with !CONFIG_MMU. Platform vm_get_page_prot() > is required only with CONFIG_MMU enabled, otherwise there is a generic fallback > stub in include/linux/mm.h > > https://lore.kernel.org/all/20220705221411.3381797-1-jcmvbkbc@gmail.com/ [xtensa] > https://lore.kernel.org/all/20220706054002.1936820-1-anshuman.khandual@arm.com/ [sh] > > It does not seem CONFIG_MMU can be disabled on other platforms thus exposing a build > failure. But just to be on the safer side, should all vm_get_page_prot() be wrapped > around with #ifdef CONFIG_MMU ? In that case will resend the series with above build > fixes folded back in as well. Please do suggest. Thank you. > As far as I can see in Kconfig, CONFIG_MMU is user selectable on the following architectures: - ARM - M68K - RISCV - SH And is disabled by default on XTENSA. So I guess ARM, M68K, RISCV, SH and XTENSA should have it wrapped around #ifdef CONFIG_MMU , or have it in a file that is not built when CONFIG_MMU is not set. Christophe
On Wed, Jul 6, 2022 at 8:33 AM Christophe Leroy <christophe.leroy@csgroup.eu> wrote: > As far as I can see in Kconfig, CONFIG_MMU is user selectable on the > following architectures: > - ARM > - M68K > - RISCV > - SH > > And is disabled by default on XTENSA. Right, the list is complete, though it's also default-enabled for every CPU core on xtensa, and you can only disable it for "custom" CPU cores. Arnd
On 7/6/22 12:34, Arnd Bergmann wrote: > On Wed, Jul 6, 2022 at 8:33 AM Christophe Leroy > <christophe.leroy@csgroup.eu> wrote: > >> As far as I can see in Kconfig, CONFIG_MMU is user selectable on the >> following architectures: >> - ARM >> - M68K >> - RISCV >> - SH >> >> And is disabled by default on XTENSA. > > Right, the list is complete, though it's also default-enabled for > every CPU core on xtensa, and you can only disable it for > "custom" CPU cores. ARM, M68K, RISCV all have got them inside wrapper or files which are not built, without CONFIG_MMU. With build fixes for SH and XTENSA on linux-next, I guess we are good.