[POWERPC] Small fixes & cleanups in segment page size demotion
authorBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 8 May 2007 06:27:27 +0000 (16:27 +1000)
committerPaul Mackerras <paulus@samba.org>
Wed, 9 May 2007 06:35:00 +0000 (16:35 +1000)
The code for demoting segments to 4K had some issues, like for example,
when using _PAGE_4K_PFN flag, the first CPU to hit it would do the
demotion, but other CPUs hitting the same page wouldn't properly flush
their SLBs if mmu_ci_restriction isn't set.  There are also potential
issues with hash_preload not handling _PAGE_4K_PFN.  All of these are
non issues on current hardware but might bite us in the future.

This patch thus fixes it by:

 - Taking the test comparing the mm and current CPU context page
sizes to decide to flush SLBs out of the mmu_ci_restrictions test
since that can also be triggered by _PAGE_4K_PFN pages

 - Due to the above being done all the time, demote_segment_4k
doesn't need update the context and flush the SLB

 - demote_segment_4k can be static and doesn't need an EXPORT_SYMBOL

 - Making hash_preload ignore anything that has either _PAGE_4K_PFN
or _PAGE_NO_CACHE set, thus avoiding duplication of the complicated
logic in hash_page() (and possibly making hash_preload a little bit
faster for the normal case).

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>
arch/powerpc/mm/hash_utils_64.c

index 9b226fa7006fd51e32d497a813abce46a38766dd..71092c2f65cdd056adebb91ae02ef089bda08e39 100644 (file)
@@ -596,22 +596,18 @@ unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
  * Demote a segment to using 4k pages.
  * For now this makes the whole process use 4k pages.
  */
-void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
-{
 #ifdef CONFIG_PPC_64K_PAGES
+static void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
+{
        if (mm->context.user_psize == MMU_PAGE_4K)
                return;
        mm->context.user_psize = MMU_PAGE_4K;
        mm->context.sllp = SLB_VSID_USER | mmu_psize_defs[MMU_PAGE_4K].sllp;
-       get_paca()->context = mm->context;
-       slb_flush_and_rebolt();
 #ifdef CONFIG_SPE_BASE
        spu_flush_all_slbs(mm);
 #endif
-#endif
 }
-
-EXPORT_SYMBOL_GPL(demote_segment_4k);
+#endif /* CONFIG_PPC_64K_PAGES */
 
 /* Result code is:
  *  0 - handled
@@ -711,40 +707,42 @@ int hash_page(unsigned long ea, unsigned long access, unsigned long trap)
                psize = MMU_PAGE_4K;
        }
 
-       if (mmu_ci_restrictions) {
-               /* If this PTE is non-cacheable, switch to 4k */
-               if (psize == MMU_PAGE_64K &&
-                   (pte_val(*ptep) & _PAGE_NO_CACHE)) {
-                       if (user_region) {
-                               demote_segment_4k(mm, ea);
-                               psize = MMU_PAGE_4K;
-                       } else if (ea < VMALLOC_END) {
-                               /*
-                                * some driver did a non-cacheable mapping
-                                * in vmalloc space, so switch vmalloc
-                                * to 4k pages
-                                */
-                               printk(KERN_ALERT "Reducing vmalloc segment "
-                                      "to 4kB pages because of "
-                                      "non-cacheable mapping\n");
-                               psize = mmu_vmalloc_psize = MMU_PAGE_4K;
-                       }
+       /* If this PTE is non-cacheable and we have restrictions on
+        * using non cacheable large pages, then we switch to 4k
+        */
+       if (mmu_ci_restrictions && psize == MMU_PAGE_64K &&
+           (pte_val(*ptep) & _PAGE_NO_CACHE)) {
+               if (user_region) {
+                       demote_segment_4k(mm, ea);
+                       psize = MMU_PAGE_4K;
+               } else if (ea < VMALLOC_END) {
+                       /*
+                        * some driver did a non-cacheable mapping
+                        * in vmalloc space, so switch vmalloc
+                        * to 4k pages
+                        */
+                       printk(KERN_ALERT "Reducing vmalloc segment "
+                              "to 4kB pages because of "
+                              "non-cacheable mapping\n");
+                       psize = mmu_vmalloc_psize = MMU_PAGE_4K;
 #ifdef CONFIG_SPE_BASE
                        spu_flush_all_slbs(mm);
 #endif
                }
-               if (user_region) {
-                       if (psize != get_paca()->context.user_psize) {
-                               get_paca()->context = mm->context;
-                               slb_flush_and_rebolt();
-                       }
-               } else if (get_paca()->vmalloc_sllp !=
-                          mmu_psize_defs[mmu_vmalloc_psize].sllp) {
-                       get_paca()->vmalloc_sllp =
-                               mmu_psize_defs[mmu_vmalloc_psize].sllp;
+       }
+       if (user_region) {
+               if (psize != get_paca()->context.user_psize) {
+                       get_paca()->context.user_psize =
+                               mm->context.user_psize;
                        slb_flush_and_rebolt();
                }
+       } else if (get_paca()->vmalloc_sllp !=
+                  mmu_psize_defs[mmu_vmalloc_psize].sllp) {
+               get_paca()->vmalloc_sllp =
+                       mmu_psize_defs[mmu_vmalloc_psize].sllp;
+               slb_flush_and_rebolt();
        }
+
        if (psize == MMU_PAGE_64K)
                rc = __hash_page_64K(ea, access, vsid, ptep, trap, local);
        else
@@ -780,13 +778,26 @@ void hash_preload(struct mm_struct *mm, unsigned long ea,
        DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
                " trap=%lx\n", mm, mm->pgd, ea, access, trap);
 
-       /* Get PTE, VSID, access mask */
+       /* Get Linux PTE if available */
        pgdir = mm->pgd;
        if (pgdir == NULL)
                return;
        ptep = find_linux_pte(pgdir, ea);
        if (!ptep)
                return;
+
+#ifdef CONFIG_PPC_64K_PAGES
+       /* If either _PAGE_4K_PFN or _PAGE_NO_CACHE is set (and we are on
+        * a 64K kernel), then we don't preload, hash_page() will take
+        * care of it once we actually try to access the page.
+        * That way we don't have to duplicate all of the logic for segment
+        * page size demotion here
+        */
+       if (pte_val(*ptep) & (_PAGE_4K_PFN | _PAGE_NO_CACHE))
+               return;
+#endif /* CONFIG_PPC_64K_PAGES */
+
+       /* Get VSID */
        vsid = get_vsid(mm->context.id, ea);
 
        /* Hash it in */
@@ -797,12 +808,6 @@ void hash_preload(struct mm_struct *mm, unsigned long ea,
 #ifndef CONFIG_PPC_64K_PAGES
        __hash_page_4K(ea, access, vsid, ptep, trap, local);
 #else
-       if (mmu_ci_restrictions) {
-               /* If this PTE is non-cacheable, switch to 4k */
-               if (mm->context.user_psize == MMU_PAGE_64K &&
-                   (pte_val(*ptep) & _PAGE_NO_CACHE))
-                       demote_segment_4k(mm, ea);
-       }
        if (mm->context.user_psize == MMU_PAGE_64K)
                __hash_page_64K(ea, access, vsid, ptep, trap, local);
        else