linux_dsm_epyc7002/arch/x86/mm/pgtable.c

448 lines
10 KiB
C
Raw Normal View History

#include <linux/mm.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 15:04:11 +07:00
#include <linux/gfp.h>
#include <asm/pgalloc.h>
#include <asm/pgtable.h>
#include <asm/tlb.h>
#include <asm/fixmap.h>
#define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
x86, mm: Allow highmem user page tables to be disabled at boot time Distros generally (I looked at Debian, RHEL5 and SLES11) seem to enable CONFIG_HIGHPTE for any x86 configuration which has highmem enabled. This means that the overhead applies even to machines which have a fairly modest amount of high memory and which therefore do not really benefit from allocating PTEs in high memory but still pay the price of the additional mapping operations. Running kernbench on a 4G box I found that with CONFIG_HIGHPTE=y but no actual highptes being allocated there was a reduction in system time used from 59.737s to 55.9s. With CONFIG_HIGHPTE=y and highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 175.396 (0.238914) User Time 515.983 (5.85019) System Time 59.737 (1.26727) Percent CPU 263.8 (71.6796) Context Switches 39989.7 (4672.64) Sleeps 42617.7 (246.307) With CONFIG_HIGHPTE=y but with no highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 174.278 (0.831968) User Time 515.659 (6.07012) System Time 55.9 (1.07799) Percent CPU 263.8 (71.266) Context Switches 39929.6 (4485.13) Sleeps 42583.7 (373.039) This patch allows the user to control the allocation of PTEs in highmem from the command line ("userpte=nohigh") but retains the status-quo as the default. It is possible that some simple heuristic could be developed which allows auto-tuning of this option however I don't have a sufficiently large machine available to me to perform any particularly meaningful experiments. We could probably handwave up an argument for a threshold at 16G of total RAM. Assuming 768M of lowmem we have 196608 potential lowmem PTE pages. Each page can map 2M of RAM in a PAE-enabled configuration, meaning a maximum of 384G of RAM could potentially be mapped using lowmem PTEs. Even allowing generous factor of 10 to account for other required lowmem allocations, generous slop to account for page sharing (which reduces the total amount of RAM mappable by a given number of PT pages) and other innacuracies in the estimations it would seem that even a 32G machine would not have a particularly pressing need for highmem PTEs. I think 32G could be considered to be at the upper bound of what might be sensible on a 32 bit machine (although I think in practice 64G is still supported). It's seems questionable if HIGHPTE is even a win for any amount of RAM you would sensibly run a 32 bit kernel on rather than going 64 bit. Signed-off-by: Ian Campbell <ian.campbell@citrix.com> LKML-Reference: <1266403090-20162-1-git-send-email-ian.campbell@citrix.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2010-02-17 17:38:10 +07:00
#ifdef CONFIG_HIGHPTE
#define PGALLOC_USER_GFP __GFP_HIGHMEM
#else
#define PGALLOC_USER_GFP 0
#endif
gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
{
return (pte_t *)__get_free_page(PGALLOC_GFP);
}
pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
{
struct page *pte;
x86, mm: Allow highmem user page tables to be disabled at boot time Distros generally (I looked at Debian, RHEL5 and SLES11) seem to enable CONFIG_HIGHPTE for any x86 configuration which has highmem enabled. This means that the overhead applies even to machines which have a fairly modest amount of high memory and which therefore do not really benefit from allocating PTEs in high memory but still pay the price of the additional mapping operations. Running kernbench on a 4G box I found that with CONFIG_HIGHPTE=y but no actual highptes being allocated there was a reduction in system time used from 59.737s to 55.9s. With CONFIG_HIGHPTE=y and highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 175.396 (0.238914) User Time 515.983 (5.85019) System Time 59.737 (1.26727) Percent CPU 263.8 (71.6796) Context Switches 39989.7 (4672.64) Sleeps 42617.7 (246.307) With CONFIG_HIGHPTE=y but with no highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 174.278 (0.831968) User Time 515.659 (6.07012) System Time 55.9 (1.07799) Percent CPU 263.8 (71.266) Context Switches 39929.6 (4485.13) Sleeps 42583.7 (373.039) This patch allows the user to control the allocation of PTEs in highmem from the command line ("userpte=nohigh") but retains the status-quo as the default. It is possible that some simple heuristic could be developed which allows auto-tuning of this option however I don't have a sufficiently large machine available to me to perform any particularly meaningful experiments. We could probably handwave up an argument for a threshold at 16G of total RAM. Assuming 768M of lowmem we have 196608 potential lowmem PTE pages. Each page can map 2M of RAM in a PAE-enabled configuration, meaning a maximum of 384G of RAM could potentially be mapped using lowmem PTEs. Even allowing generous factor of 10 to account for other required lowmem allocations, generous slop to account for page sharing (which reduces the total amount of RAM mappable by a given number of PT pages) and other innacuracies in the estimations it would seem that even a 32G machine would not have a particularly pressing need for highmem PTEs. I think 32G could be considered to be at the upper bound of what might be sensible on a 32 bit machine (although I think in practice 64G is still supported). It's seems questionable if HIGHPTE is even a win for any amount of RAM you would sensibly run a 32 bit kernel on rather than going 64 bit. Signed-off-by: Ian Campbell <ian.campbell@citrix.com> LKML-Reference: <1266403090-20162-1-git-send-email-ian.campbell@citrix.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2010-02-17 17:38:10 +07:00
pte = alloc_pages(__userpte_alloc_gfp, 0);
if (pte)
pgtable_page_ctor(pte);
return pte;
}
x86, mm: Allow highmem user page tables to be disabled at boot time Distros generally (I looked at Debian, RHEL5 and SLES11) seem to enable CONFIG_HIGHPTE for any x86 configuration which has highmem enabled. This means that the overhead applies even to machines which have a fairly modest amount of high memory and which therefore do not really benefit from allocating PTEs in high memory but still pay the price of the additional mapping operations. Running kernbench on a 4G box I found that with CONFIG_HIGHPTE=y but no actual highptes being allocated there was a reduction in system time used from 59.737s to 55.9s. With CONFIG_HIGHPTE=y and highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 175.396 (0.238914) User Time 515.983 (5.85019) System Time 59.737 (1.26727) Percent CPU 263.8 (71.6796) Context Switches 39989.7 (4672.64) Sleeps 42617.7 (246.307) With CONFIG_HIGHPTE=y but with no highmem PTEs being allocated: Average Optimal load -j 4 Run (std deviation): Elapsed Time 174.278 (0.831968) User Time 515.659 (6.07012) System Time 55.9 (1.07799) Percent CPU 263.8 (71.266) Context Switches 39929.6 (4485.13) Sleeps 42583.7 (373.039) This patch allows the user to control the allocation of PTEs in highmem from the command line ("userpte=nohigh") but retains the status-quo as the default. It is possible that some simple heuristic could be developed which allows auto-tuning of this option however I don't have a sufficiently large machine available to me to perform any particularly meaningful experiments. We could probably handwave up an argument for a threshold at 16G of total RAM. Assuming 768M of lowmem we have 196608 potential lowmem PTE pages. Each page can map 2M of RAM in a PAE-enabled configuration, meaning a maximum of 384G of RAM could potentially be mapped using lowmem PTEs. Even allowing generous factor of 10 to account for other required lowmem allocations, generous slop to account for page sharing (which reduces the total amount of RAM mappable by a given number of PT pages) and other innacuracies in the estimations it would seem that even a 32G machine would not have a particularly pressing need for highmem PTEs. I think 32G could be considered to be at the upper bound of what might be sensible on a 32 bit machine (although I think in practice 64G is still supported). It's seems questionable if HIGHPTE is even a win for any amount of RAM you would sensibly run a 32 bit kernel on rather than going 64 bit. Signed-off-by: Ian Campbell <ian.campbell@citrix.com> LKML-Reference: <1266403090-20162-1-git-send-email-ian.campbell@citrix.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2010-02-17 17:38:10 +07:00
static int __init setup_userpte(char *arg)
{
if (!arg)
return -EINVAL;
/*
* "userpte=nohigh" disables allocation of user pagetables in
* high memory.
*/
if (strcmp(arg, "nohigh") == 0)
__userpte_alloc_gfp &= ~__GFP_HIGHMEM;
else
return -EINVAL;
return 0;
}
early_param("userpte", setup_userpte);
void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
{
pgtable_page_dtor(pte);
paravirt_release_pte(page_to_pfn(pte));
tlb_remove_page(tlb, pte);
}
#if PAGETABLE_LEVELS > 2
void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
{
paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
tlb_remove_page(tlb, virt_to_page(pmd));
}
#if PAGETABLE_LEVELS > 3
void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
{
paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
tlb_remove_page(tlb, virt_to_page(pud));
}
#endif /* PAGETABLE_LEVELS > 3 */
#endif /* PAGETABLE_LEVELS > 2 */
static inline void pgd_list_add(pgd_t *pgd)
{
struct page *page = virt_to_page(pgd);
list_add(&page->lru, &pgd_list);
}
static inline void pgd_list_del(pgd_t *pgd)
{
struct page *page = virt_to_page(pgd);
list_del(&page->lru);
}
#define UNSHARED_PTRS_PER_PGD \
(SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
{
BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
virt_to_page(pgd)->index = (pgoff_t)mm;
}
struct mm_struct *pgd_page_get_mm(struct page *page)
{
return (struct mm_struct *)page->index;
}
static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
{
/* If the pgd points to a shared pagetable level (either the
ptes in non-PAE, or shared PMD in PAE), then just copy the
references from swapper_pg_dir. */
if (PAGETABLE_LEVELS == 2 ||
(PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
PAGETABLE_LEVELS == 4) {
clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
swapper_pg_dir + KERNEL_PGD_BOUNDARY,
KERNEL_PGD_PTRS);
}
/* list required to sync kernel mapping updates */
if (!SHARED_KERNEL_PMD) {
pgd_set_mm(pgd, mm);
pgd_list_add(pgd);
}
}
static void pgd_dtor(pgd_t *pgd)
{
if (SHARED_KERNEL_PMD)
return;
spin_lock(&pgd_lock);
pgd_list_del(pgd);
spin_unlock(&pgd_lock);
}
/*
* List of all pgd's needed for non-PAE so it can invalidate entries
* in both cached and uncached pgd's; not needed for PAE since the
* kernel pmd is shared. If PAE were not to share the pmd a similar
* tactic would be needed. This is essentially codepath-based locking
* against pageattr.c; it is the unique case in which a valid change
* of kernel pagetables can't be lazily synchronized by vmalloc faults.
* vmalloc faults work because attached pagetables are never freed.
* -- wli
*/
#ifdef CONFIG_X86_PAE
/*
* In PAE mode, we need to do a cr3 reload (=tlb flush) when
* updating the top-level pagetable entries to guarantee the
* processor notices the update. Since this is expensive, and
* all 4 top-level entries are used almost immediately in a
* new process's life, we just pre-populate them here.
*
* Also, if we're in a paravirt environment where the kernel pmd is
* not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
* and initialize the kernel pmds here.
*/
#define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
{
paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
/* Note: almost everything apart from _PAGE_PRESENT is
reserved at the pmd (PDPT) level. */
set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
/*
* According to Intel App note "TLBs, Paging-Structure Caches,
* and Their Invalidation", April 2007, document 317080-001,
* section 8.1: in PAE mode we explicitly have to flush the
* TLB via cr3 if the top-level pgd is changed...
*/
flush_tlb_mm(mm);
}
#else /* !CONFIG_X86_PAE */
/* No need to prepopulate any pagetable entries in non-PAE modes. */
#define PREALLOCATED_PMDS 0
#endif /* CONFIG_X86_PAE */
static void free_pmds(pmd_t *pmds[])
{
int i;
for(i = 0; i < PREALLOCATED_PMDS; i++)
if (pmds[i])
free_page((unsigned long)pmds[i]);
}
static int preallocate_pmds(pmd_t *pmds[])
{
int i;
bool failed = false;
for(i = 0; i < PREALLOCATED_PMDS; i++) {
pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
if (pmd == NULL)
failed = true;
pmds[i] = pmd;
}
if (failed) {
free_pmds(pmds);
return -ENOMEM;
}
return 0;
}
/*
* Mop up any pmd pages which may still be attached to the pgd.
* Normally they will be freed by munmap/exit_mmap, but any pmd we
* preallocate which never got a corresponding vma will need to be
* freed manually.
*/
static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
{
int i;
for(i = 0; i < PREALLOCATED_PMDS; i++) {
pgd_t pgd = pgdp[i];
if (pgd_val(pgd) != 0) {
pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
pgdp[i] = native_make_pgd(0);
paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
pmd_free(mm, pmd);
}
}
}
static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
{
pud_t *pud;
unsigned long addr;
int i;
if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
return;
pud = pud_offset(pgd, 0);
for (addr = i = 0; i < PREALLOCATED_PMDS;
i++, pud++, addr += PUD_SIZE) {
pmd_t *pmd = pmds[i];
if (i >= KERNEL_PGD_BOUNDARY)
memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
sizeof(pmd_t) * PTRS_PER_PMD);
pud_populate(mm, pud, pmd);
}
}
pgd_t *pgd_alloc(struct mm_struct *mm)
{
pgd_t *pgd;
pmd_t *pmds[PREALLOCATED_PMDS];
pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
if (pgd == NULL)
goto out;
mm->pgd = pgd;
if (preallocate_pmds(pmds) != 0)
goto out_free_pgd;
if (paravirt_pgd_alloc(mm) != 0)
goto out_free_pmds;
/*
* Make sure that pre-populating the pmds is atomic with
* respect to anything walking the pgd_list, so that they
* never see a partially populated pgd.
*/
spin_lock(&pgd_lock);
pgd_ctor(mm, pgd);
pgd_prepopulate_pmd(mm, pgd, pmds);
spin_unlock(&pgd_lock);
return pgd;
out_free_pmds:
free_pmds(pmds);
out_free_pgd:
free_page((unsigned long)pgd);
out:
return NULL;
}
void pgd_free(struct mm_struct *mm, pgd_t *pgd)
{
pgd_mop_up_pmds(mm, pgd);
pgd_dtor(pgd);
paravirt_pgd_free(mm, pgd);
free_page((unsigned long)pgd);
}
int ptep_set_access_flags(struct vm_area_struct *vma,
unsigned long address, pte_t *ptep,
pte_t entry, int dirty)
{
int changed = !pte_same(*ptep, entry);
if (changed && dirty) {
*ptep = entry;
pte_update_defer(vma->vm_mm, address, ptep);
flush_tlb_page(vma, address);
}
return changed;
}
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
int pmdp_set_access_flags(struct vm_area_struct *vma,
unsigned long address, pmd_t *pmdp,
pmd_t entry, int dirty)
{
int changed = !pmd_same(*pmdp, entry);
VM_BUG_ON(address & ~HPAGE_PMD_MASK);
if (changed && dirty) {
*pmdp = entry;
pmd_update_defer(vma->vm_mm, address, pmdp);
flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
}
return changed;
}
#endif
int ptep_test_and_clear_young(struct vm_area_struct *vma,
unsigned long addr, pte_t *ptep)
{
int ret = 0;
if (pte_young(*ptep))
ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
(unsigned long *) &ptep->pte);
if (ret)
pte_update(vma->vm_mm, addr, ptep);
return ret;
}
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
int pmdp_test_and_clear_young(struct vm_area_struct *vma,
unsigned long addr, pmd_t *pmdp)
{
int ret = 0;
if (pmd_young(*pmdp))
ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
(unsigned long *)pmdp);
if (ret)
pmd_update(vma->vm_mm, addr, pmdp);
return ret;
}
#endif
int ptep_clear_flush_young(struct vm_area_struct *vma,
unsigned long address, pte_t *ptep)
{
int young;
young = ptep_test_and_clear_young(vma, address, ptep);
if (young)
flush_tlb_page(vma, address);
return young;
}
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
int pmdp_clear_flush_young(struct vm_area_struct *vma,
unsigned long address, pmd_t *pmdp)
{
int young;
VM_BUG_ON(address & ~HPAGE_PMD_MASK);
young = pmdp_test_and_clear_young(vma, address, pmdp);
if (young)
flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
return young;
}
void pmdp_splitting_flush(struct vm_area_struct *vma,
unsigned long address, pmd_t *pmdp)
{
int set;
VM_BUG_ON(address & ~HPAGE_PMD_MASK);
set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
(unsigned long *)pmdp);
if (set) {
pmd_update(vma->vm_mm, address, pmdp);
/* need tlb flush only to serialize against gup-fast */
flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
}
}
#endif
/**
* reserve_top_address - reserves a hole in the top of kernel address space
* @reserve - size of hole to reserve
*
* Can be used to relocate the fixmap area and poke a hole in the top
* of kernel address space to make room for a hypervisor.
*/
void __init reserve_top_address(unsigned long reserve)
{
#ifdef CONFIG_X86_32
BUG_ON(fixmaps_set > 0);
printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
(int)-reserve);
__FIXADDR_TOP = -reserve - PAGE_SIZE;
#endif
}
int fixmaps_set;
void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
{
unsigned long address = __fix_to_virt(idx);
if (idx >= __end_of_fixed_addresses) {
BUG();
return;
}
set_pte_vaddr(address, pte);
fixmaps_set++;
}
void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
pgprot_t flags)
{
__native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
}