| Message ID | 20221221174454.1085130-1-urezki@gmail.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v2,1/3] mm: vmalloc: Avoid of calling __find_vmap_area() twise in __vunmap() | expand |
Some pedantic grammar/spelling stuff:- (I know it can be a little annoying to get grammatical suggestions so I do hope that it isn't too irritating!) For the Subject line:- 'mm: vmalloc: Avoid of calling __find_vmap_area() twise in __vunmap()' -> 'mm: vmalloc: Avoid calling __find_vmap_area() twice in __vunmap()' On Wed, Dec 21, 2022 at 06:44:52PM +0100, Uladzislau Rezki (Sony) wrote: > Currently __vunmap() path calls __find_vmap_area() two times. One on > entry to check that area exists, second time inside remove_vm_area() > function that also performs a new search of VA. Perhaps slightly tweak to:- "Currently the __vunmap() path calls __find_vmap_area() twice. Once on entry to check that the area exists, then inside the remove_vm_area() function which also performs a new search for the VA." > > In order to improvie it from a performance point of view we split > remove_vm_area() into two new parts: > - find_unlink_vmap_area() that does a search and unlink from tree; > - __remove_vm_area() that does a removing but without searching. 'that does a removing but without searching' reads better I think as 'that removes without searching'. > > In this case there is no any functional change for remove_vm_area() > whereas vm_remove_mappings(), where a second search happens, switches > to the __remove_vm_area() variant where already detached VA is passed > as a parameter, so there is no need to find it again. > 'where already detached VA' -> 'where the already detached VA' as a minor nit here! > Performance wise, i use test_vmalloc.sh with 32 threads doing alloc > free on a 64-CPUs-x86_64-box: > > perf without this patch: > - 31.41% 0.50% vmalloc_test/10 [kernel.vmlinux] [k] __vunmap > - 30.92% __vunmap > - 17.67% _raw_spin_lock > native_queued_spin_lock_slowpath > - 12.33% remove_vm_area > - 11.79% free_vmap_area_noflush > - 11.18% _raw_spin_lock > native_queued_spin_lock_slowpath > 0.76% free_unref_page > > perf with this patch: > - 11.35% 0.13% vmalloc_test/14 [kernel.vmlinux] [k] __vunmap > - 11.23% __vunmap > - 8.28% find_unlink_vmap_area > - 7.95% _raw_spin_lock > 7.44% native_queued_spin_lock_slowpath > - 1.93% free_vmap_area_noflush > - 0.56% _raw_spin_lock > 0.53% native_queued_spin_lock_slowpath > 0.60% __vunmap_range_noflush > > __vunmap() consumes around ~20% less CPU cycles on this test. Very nice, amazing work! > > Reported-by: Roman Gushchin <roman.gushchin@linux.dev> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> > --- > mm/vmalloc.c | 66 +++++++++++++++++++++++++++++++++------------------- > 1 file changed, 42 insertions(+), 24 deletions(-) > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > index 9e30f0b39203..28030d2441f1 100644 > --- a/mm/vmalloc.c > +++ b/mm/vmalloc.c > @@ -1825,9 +1825,11 @@ static void free_vmap_area_noflush(struct vmap_area *va) > unsigned long va_start = va->va_start; > unsigned long nr_lazy; > > - spin_lock(&vmap_area_lock); > - unlink_va(va, &vmap_area_root); > - spin_unlock(&vmap_area_lock); > + if (!list_empty(&va->list)) { > + spin_lock(&vmap_area_lock); > + unlink_va(va, &vmap_area_root); > + spin_unlock(&vmap_area_lock); > + } Do we want to do the same in free_vmap_area()? > > nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> > PAGE_SHIFT, &vmap_lazy_nr); > @@ -1871,6 +1873,19 @@ struct vmap_area *find_vmap_area(unsigned long addr) > return va; > } > > +static struct vmap_area *find_unlink_vmap_area(unsigned long addr) > +{ > + struct vmap_area *va; > + > + spin_lock(&vmap_area_lock); > + va = __find_vmap_area(addr, &vmap_area_root); > + if (va) > + unlink_va(va, &vmap_area_root); > + spin_unlock(&vmap_area_lock); > + > + return va; > +} > + > /*** Per cpu kva allocator ***/ > > /* > @@ -2591,6 +2606,20 @@ struct vm_struct *find_vm_area(const void *addr) > return va->vm; > } > > +static struct vm_struct *__remove_vm_area(struct vmap_area *va) > +{ > + struct vm_struct *vm; > + > + if (!va || !va->vm) > + return NULL; > + > + vm = va->vm; > + kasan_free_module_shadow(vm); > + free_unmap_vmap_area(va); > + > + return vm; > +} > + > /** > * remove_vm_area - find and remove a continuous kernel virtual area > * @addr: base address > @@ -2607,22 +2636,8 @@ struct vm_struct *remove_vm_area(const void *addr) > > might_sleep(); > > - spin_lock(&vmap_area_lock); > - va = __find_vmap_area((unsigned long)addr, &vmap_area_root); > - if (va && va->vm) { > - struct vm_struct *vm = va->vm; > - > - va->vm = NULL; > - spin_unlock(&vmap_area_lock); > - > - kasan_free_module_shadow(vm); > - free_unmap_vmap_area(va); > - > - return vm; > - } > - > - spin_unlock(&vmap_area_lock); > - return NULL; > + va = find_unlink_vmap_area((unsigned long) addr); > + return __remove_vm_area(va); > } Really nice separation of concerns and cleanup. > > static inline void set_area_direct_map(const struct vm_struct *area, > @@ -2637,15 +2652,16 @@ static inline void set_area_direct_map(const struct vm_struct *area, > } > > /* Handle removing and resetting vm mappings related to the vm_struct. */ > -static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > +static void vm_remove_mappings(struct vmap_area *va, int deallocate_pages) Perhaps rename this to va_remove_mappings() or vmap_area_remove_mappings() since it is now explicitly accepting a vmap_area rather than vm_struct? > { > + struct vm_struct *area = va->vm; > unsigned long start = ULONG_MAX, end = 0; > unsigned int page_order = vm_area_page_order(area); > int flush_reset = area->flags & VM_FLUSH_RESET_PERMS; > int flush_dmap = 0; > int i; > > - remove_vm_area(area->addr); > + __remove_vm_area(va); > > /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */ > if (!flush_reset) > @@ -2690,6 +2706,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > static void __vunmap(const void *addr, int deallocate_pages) > { > struct vm_struct *area; Feels like it's getting a bit confusing with 'va' representing vmap_area and 'area' which represents... vm_struct (this file has a bunch of naming inconsistencies like this actually), perhaps rename this to 'vm'? > + struct vmap_area *va; > > if (!addr) > return; > @@ -2698,19 +2715,20 @@ static void __vunmap(const void *addr, int deallocate_pages) > addr)) > return; > > - area = find_vm_area(addr); > - if (unlikely(!area)) { > + va = find_unlink_vmap_area((unsigned long)addr); > + if (unlikely(!va)) { > WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", > addr); > return; > } > > + area = va->vm; > debug_check_no_locks_freed(area->addr, get_vm_area_size(area)); > debug_check_no_obj_freed(area->addr, get_vm_area_size(area)); > > kasan_poison_vmalloc(area->addr, get_vm_area_size(area)); > > - vm_remove_mappings(area, deallocate_pages); > + vm_remove_mappings(va, deallocate_pages); > > if (deallocate_pages) { > int i; > -- > 2.30.2 > Other than some pendatic points about grammar/naming this looks really good!
A sorry. I need more coffee, this is not the cover letter, but the patch that introduceѕ find_unlink_vmap_area. > - spin_lock(&vmap_area_lock); > - unlink_va(va, &vmap_area_root); > - spin_unlock(&vmap_area_lock); > + if (!list_empty(&va->list)) { > + spin_lock(&vmap_area_lock); > + unlink_va(va, &vmap_area_root); > + spin_unlock(&vmap_area_lock); > + } As mentioned before, I'd much rather move this into the callers. > + va = find_unlink_vmap_area((unsigned long) addr); > + return __remove_vm_area(va); This can drop the va local variable now.
On 12/21/22 at 06:44pm, Uladzislau Rezki (Sony) wrote: > Currently __vunmap() path calls __find_vmap_area() two times. One on > entry to check that area exists, second time inside remove_vm_area() > function that also performs a new search of VA. > > In order to improvie it from a performance point of view we split > remove_vm_area() into two new parts: > - find_unlink_vmap_area() that does a search and unlink from tree; > - __remove_vm_area() that does a removing but without searching. > > In this case there is no any functional change for remove_vm_area() > whereas vm_remove_mappings(), where a second search happens, switches > to the __remove_vm_area() variant where already detached VA is passed > as a parameter, so there is no need to find it again. I like this patch. This takes off the va->vm clearning too. Finally I don't need to worry about the va->flags clearing during unmapping when reading out vmap_block areas. > > Performance wise, i use test_vmalloc.sh with 32 threads doing alloc > free on a 64-CPUs-x86_64-box: > > perf without this patch: > - 31.41% 0.50% vmalloc_test/10 [kernel.vmlinux] [k] __vunmap > - 30.92% __vunmap > - 17.67% _raw_spin_lock > native_queued_spin_lock_slowpath > - 12.33% remove_vm_area > - 11.79% free_vmap_area_noflush > - 11.18% _raw_spin_lock > native_queued_spin_lock_slowpath > 0.76% free_unref_page > > perf with this patch: > - 11.35% 0.13% vmalloc_test/14 [kernel.vmlinux] [k] __vunmap > - 11.23% __vunmap > - 8.28% find_unlink_vmap_area > - 7.95% _raw_spin_lock > 7.44% native_queued_spin_lock_slowpath > - 1.93% free_vmap_area_noflush > - 0.56% _raw_spin_lock > 0.53% native_queued_spin_lock_slowpath > 0.60% __vunmap_range_noflush > > __vunmap() consumes around ~20% less CPU cycles on this test. > > Reported-by: Roman Gushchin <roman.gushchin@linux.dev> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> > --- > mm/vmalloc.c | 66 +++++++++++++++++++++++++++++++++------------------- > 1 file changed, 42 insertions(+), 24 deletions(-) > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > index 9e30f0b39203..28030d2441f1 100644 > --- a/mm/vmalloc.c > +++ b/mm/vmalloc.c > @@ -1825,9 +1825,11 @@ static void free_vmap_area_noflush(struct vmap_area *va) > unsigned long va_start = va->va_start; > unsigned long nr_lazy; > > - spin_lock(&vmap_area_lock); > - unlink_va(va, &vmap_area_root); > - spin_unlock(&vmap_area_lock); > + if (!list_empty(&va->list)) { > + spin_lock(&vmap_area_lock); > + unlink_va(va, &vmap_area_root); > + spin_unlock(&vmap_area_lock); > + } > > nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> > PAGE_SHIFT, &vmap_lazy_nr); > @@ -1871,6 +1873,19 @@ struct vmap_area *find_vmap_area(unsigned long addr) > return va; > } > > +static struct vmap_area *find_unlink_vmap_area(unsigned long addr) > +{ > + struct vmap_area *va; > + > + spin_lock(&vmap_area_lock); > + va = __find_vmap_area(addr, &vmap_area_root); > + if (va) > + unlink_va(va, &vmap_area_root); > + spin_unlock(&vmap_area_lock); > + > + return va; > +} > + > /*** Per cpu kva allocator ***/ > > /* > @@ -2591,6 +2606,20 @@ struct vm_struct *find_vm_area(const void *addr) > return va->vm; > } > > +static struct vm_struct *__remove_vm_area(struct vmap_area *va) > +{ > + struct vm_struct *vm; > + > + if (!va || !va->vm) > + return NULL; > + > + vm = va->vm; > + kasan_free_module_shadow(vm); > + free_unmap_vmap_area(va); > + > + return vm; > +} > + > /** > * remove_vm_area - find and remove a continuous kernel virtual area > * @addr: base address > @@ -2607,22 +2636,8 @@ struct vm_struct *remove_vm_area(const void *addr) > > might_sleep(); > > - spin_lock(&vmap_area_lock); > - va = __find_vmap_area((unsigned long)addr, &vmap_area_root); > - if (va && va->vm) { > - struct vm_struct *vm = va->vm; > - > - va->vm = NULL; > - spin_unlock(&vmap_area_lock); > - > - kasan_free_module_shadow(vm); > - free_unmap_vmap_area(va); > - > - return vm; > - } > - > - spin_unlock(&vmap_area_lock); > - return NULL; > + va = find_unlink_vmap_area((unsigned long) addr); > + return __remove_vm_area(va); > } > > static inline void set_area_direct_map(const struct vm_struct *area, > @@ -2637,15 +2652,16 @@ static inline void set_area_direct_map(const struct vm_struct *area, > } > > /* Handle removing and resetting vm mappings related to the vm_struct. */ > -static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > +static void vm_remove_mappings(struct vmap_area *va, int deallocate_pages) > { > + struct vm_struct *area = va->vm; > unsigned long start = ULONG_MAX, end = 0; > unsigned int page_order = vm_area_page_order(area); > int flush_reset = area->flags & VM_FLUSH_RESET_PERMS; > int flush_dmap = 0; > int i; > > - remove_vm_area(area->addr); > + __remove_vm_area(va); > > /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */ > if (!flush_reset) > @@ -2690,6 +2706,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > static void __vunmap(const void *addr, int deallocate_pages) > { > struct vm_struct *area; > + struct vmap_area *va; > > if (!addr) > return; > @@ -2698,19 +2715,20 @@ static void __vunmap(const void *addr, int deallocate_pages) > addr)) > return; > > - area = find_vm_area(addr); > - if (unlikely(!area)) { > + va = find_unlink_vmap_area((unsigned long)addr); > + if (unlikely(!va)) { > WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", > addr); > return; > } > > + area = va->vm; > debug_check_no_locks_freed(area->addr, get_vm_area_size(area)); > debug_check_no_obj_freed(area->addr, get_vm_area_size(area)); > > kasan_poison_vmalloc(area->addr, get_vm_area_size(area)); > > - vm_remove_mappings(area, deallocate_pages); > + vm_remove_mappings(va, deallocate_pages); > > if (deallocate_pages) { > int i; > -- > 2.30.2 >
> Some pedantic grammar/spelling stuff:- > > (I know it can be a little annoying to get grammatical suggestions so I do hope > that it isn't too irritating!) > It is absolutely OK :) > > For the Subject line:- > 'mm: vmalloc: Avoid of calling __find_vmap_area() twise in __vunmap()' -> > 'mm: vmalloc: Avoid calling __find_vmap_area() twice in __vunmap()' > Will fix in the v3. > On Wed, Dec 21, 2022 at 06:44:52PM +0100, Uladzislau Rezki (Sony) wrote: > > Currently __vunmap() path calls __find_vmap_area() two times. One on > > entry to check that area exists, second time inside remove_vm_area() > > function that also performs a new search of VA. > > Perhaps slightly tweak to:- > > "Currently the __vunmap() path calls __find_vmap_area() twice. Once on entry > to check that the area exists, then inside the remove_vm_area() function > which also performs a new search for the VA." > Will fix in the v3. > > > > In order to improvie it from a performance point of view we split > > remove_vm_area() into two new parts: > > - find_unlink_vmap_area() that does a search and unlink from tree; > > - __remove_vm_area() that does a removing but without searching. > > 'that does a removing but without searching' reads better I think as > 'that removes without searching'. > Will fix in the v3. > > > > In this case there is no any functional change for remove_vm_area() > > whereas vm_remove_mappings(), where a second search happens, switches > > to the __remove_vm_area() variant where already detached VA is passed > > as a parameter, so there is no need to find it again. > > > > 'where already detached VA' -> 'where the already detached VA' as a minor nit > here! > Will fix it. > > Performance wise, i use test_vmalloc.sh with 32 threads doing alloc > > free on a 64-CPUs-x86_64-box: > > > > perf without this patch: > > - 31.41% 0.50% vmalloc_test/10 [kernel.vmlinux] [k] __vunmap > > - 30.92% __vunmap > > - 17.67% _raw_spin_lock > > native_queued_spin_lock_slowpath > > - 12.33% remove_vm_area > > - 11.79% free_vmap_area_noflush > > - 11.18% _raw_spin_lock > > native_queued_spin_lock_slowpath > > 0.76% free_unref_page > > > > perf with this patch: > > - 11.35% 0.13% vmalloc_test/14 [kernel.vmlinux] [k] __vunmap > > - 11.23% __vunmap > > - 8.28% find_unlink_vmap_area > > - 7.95% _raw_spin_lock > > 7.44% native_queued_spin_lock_slowpath > > - 1.93% free_vmap_area_noflush > > - 0.56% _raw_spin_lock > > 0.53% native_queued_spin_lock_slowpath > > 0.60% __vunmap_range_noflush > > > > __vunmap() consumes around ~20% less CPU cycles on this test. > > Very nice, amazing work! > Thanks! > > > > Reported-by: Roman Gushchin <roman.gushchin@linux.dev> > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> > > --- > > mm/vmalloc.c | 66 +++++++++++++++++++++++++++++++++------------------- > > 1 file changed, 42 insertions(+), 24 deletions(-) > > > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > > index 9e30f0b39203..28030d2441f1 100644 > > --- a/mm/vmalloc.c > > +++ b/mm/vmalloc.c > > @@ -1825,9 +1825,11 @@ static void free_vmap_area_noflush(struct vmap_area *va) > > unsigned long va_start = va->va_start; > > unsigned long nr_lazy; > > > > - spin_lock(&vmap_area_lock); > > - unlink_va(va, &vmap_area_root); > > - spin_unlock(&vmap_area_lock); > > + if (!list_empty(&va->list)) { > > + spin_lock(&vmap_area_lock); > > + unlink_va(va, &vmap_area_root); > > + spin_unlock(&vmap_area_lock); > > + } > > Do we want to do the same in free_vmap_area()? > The free_vmap_area() is a bit special. It only pairs with alloc_vmap_area(). There are two users and both invoke free_vmap_area() in a error path. So probably it would be good to remove it fully. But it requires some refactoring. > > > > nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> > > PAGE_SHIFT, &vmap_lazy_nr); > > @@ -1871,6 +1873,19 @@ struct vmap_area *find_vmap_area(unsigned long addr) > > return va; > > } > > > > +static struct vmap_area *find_unlink_vmap_area(unsigned long addr) > > +{ > > + struct vmap_area *va; > > + > > + spin_lock(&vmap_area_lock); > > + va = __find_vmap_area(addr, &vmap_area_root); > > + if (va) > > + unlink_va(va, &vmap_area_root); > > + spin_unlock(&vmap_area_lock); > > + > > + return va; > > +} > > + > > /*** Per cpu kva allocator ***/ > > > > /* > > @@ -2591,6 +2606,20 @@ struct vm_struct *find_vm_area(const void *addr) > > return va->vm; > > } > > > > +static struct vm_struct *__remove_vm_area(struct vmap_area *va) > > +{ > > + struct vm_struct *vm; > > + > > + if (!va || !va->vm) > > + return NULL; > > + > > + vm = va->vm; > > + kasan_free_module_shadow(vm); > > + free_unmap_vmap_area(va); > > + > > + return vm; > > +} > > + > > /** > > * remove_vm_area - find and remove a continuous kernel virtual area > > * @addr: base address > > @@ -2607,22 +2636,8 @@ struct vm_struct *remove_vm_area(const void *addr) > > > > might_sleep(); > > > > - spin_lock(&vmap_area_lock); > > - va = __find_vmap_area((unsigned long)addr, &vmap_area_root); > > - if (va && va->vm) { > > - struct vm_struct *vm = va->vm; > > - > > - va->vm = NULL; > > - spin_unlock(&vmap_area_lock); > > - > > - kasan_free_module_shadow(vm); > > - free_unmap_vmap_area(va); > > - > > - return vm; > > - } > > - > > - spin_unlock(&vmap_area_lock); > > - return NULL; > > + va = find_unlink_vmap_area((unsigned long) addr); > > + return __remove_vm_area(va); > > } > > Really nice separation of concerns and cleanup. > Thanks! > > > > static inline void set_area_direct_map(const struct vm_struct *area, > > @@ -2637,15 +2652,16 @@ static inline void set_area_direct_map(const struct vm_struct *area, > > } > > > > /* Handle removing and resetting vm mappings related to the vm_struct. */ > > -static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > > +static void vm_remove_mappings(struct vmap_area *va, int deallocate_pages) > > Perhaps rename this to va_remove_mappings() or vmap_area_remove_mappings() since > it is now explicitly accepting a vmap_area rather than vm_struct? > I agree. There is a discrepancy. I can rename it to the va_remove_mappings() if there are no any complains from others. > > { > > + struct vm_struct *area = va->vm; > > unsigned long start = ULONG_MAX, end = 0; > > unsigned int page_order = vm_area_page_order(area); > > int flush_reset = area->flags & VM_FLUSH_RESET_PERMS; > > int flush_dmap = 0; > > int i; > > > > - remove_vm_area(area->addr); > > + __remove_vm_area(va); > > > > /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */ > > if (!flush_reset) > > @@ -2690,6 +2706,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) > > static void __vunmap(const void *addr, int deallocate_pages) > > { > > struct vm_struct *area; > > Feels like it's getting a bit confusing with 'va' representing vmap_area and > 'area' which represents... vm_struct (this file has a bunch of naming > inconsistencies like this actually), perhaps rename this to 'vm'? > We can. I think it should be a separate patch-set for refactoring of things like: va, vm, area, vmap, etc :) > > + struct vmap_area *va; > > > > if (!addr) > > return; > > @@ -2698,19 +2715,20 @@ static void __vunmap(const void *addr, int deallocate_pages) > > addr)) > > return; > > > > - area = find_vm_area(addr); > > - if (unlikely(!area)) { > > + va = find_unlink_vmap_area((unsigned long)addr); > > + if (unlikely(!va)) { > > WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", > > addr); > > return; > > } > > > > + area = va->vm; > > debug_check_no_locks_freed(area->addr, get_vm_area_size(area)); > > debug_check_no_obj_freed(area->addr, get_vm_area_size(area)); > > > > kasan_poison_vmalloc(area->addr, get_vm_area_size(area)); > > > > - vm_remove_mappings(area, deallocate_pages); > > + vm_remove_mappings(va, deallocate_pages); > > > > if (deallocate_pages) { > > int i; > > -- > > 2.30.2 > > > > Other than some pendatic points about grammar/naming this looks really good! > Thank you for the review! -- Uladzislau Rezki
> A sorry. I need more coffee, this is not the cover letter, but the > patch that introduceѕ find_unlink_vmap_area. > Sorry. I should post it with a cover latter to make it less confusing. > > - spin_lock(&vmap_area_lock); > > - unlink_va(va, &vmap_area_root); > > - spin_unlock(&vmap_area_lock); > > + if (!list_empty(&va->list)) { > > + spin_lock(&vmap_area_lock); > > + unlink_va(va, &vmap_area_root); > > + spin_unlock(&vmap_area_lock); > > + } > > As mentioned before, I'd much rather move this into the callers. > Agree. There is only one caller, it is the free_vmap_block(). Will fix in the v3. > > + va = find_unlink_vmap_area((unsigned long) addr); > > + return __remove_vm_area(va); > > This can drop the va local variable now. > Do you mean like: struct vm_struct *remove_vm_area(const void *addr) { might_sleep(); return __remove_vm_area( find_unlink_vmap_area((unsigned long) addr)); } ? Thanks for review! -- Uladzislau Rezki
On Thu, Dec 22, 2022 at 07:38:14PM +0800, Baoquan He wrote: > On 12/21/22 at 06:44pm, Uladzislau Rezki (Sony) wrote: > > Currently __vunmap() path calls __find_vmap_area() two times. One on > > entry to check that area exists, second time inside remove_vm_area() > > function that also performs a new search of VA. > > > > In order to improvie it from a performance point of view we split > > remove_vm_area() into two new parts: > > - find_unlink_vmap_area() that does a search and unlink from tree; > > - __remove_vm_area() that does a removing but without searching. > > > > In this case there is no any functional change for remove_vm_area() > > whereas vm_remove_mappings(), where a second search happens, switches > > to the __remove_vm_area() variant where already detached VA is passed > > as a parameter, so there is no need to find it again. > > I like this patch. This takes off the va->vm clearning too. Finally I > don't need to worry about the va->flags clearing during unmapping > when reading out vmap_block areas. > Thanks. This patch was one of the reason to help out with the per-cpu busy areas tracking/reading of your work :) -- Uladzsislau Rezki
On Thu, Dec 22, 2022 at 03:41:29PM +0100, Uladzislau Rezki wrote:
> Do you mean like:
Yes.
On Thu, Dec 22, 2022 at 07:01:05AM -0800, Christoph Hellwig wrote: > On Thu, Dec 22, 2022 at 03:41:29PM +0100, Uladzislau Rezki wrote: > > Do you mean like: > > Yes. > OK. Will eliminate that local variable. Thanks. -- Uladzislau Rezki
diff --git a/mm/vmalloc.c b/mm/vmalloc.c index 9e30f0b39203..28030d2441f1 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -1825,9 +1825,11 @@ static void free_vmap_area_noflush(struct vmap_area *va) unsigned long va_start = va->va_start; unsigned long nr_lazy; - spin_lock(&vmap_area_lock); - unlink_va(va, &vmap_area_root); - spin_unlock(&vmap_area_lock); + if (!list_empty(&va->list)) { + spin_lock(&vmap_area_lock); + unlink_va(va, &vmap_area_root); + spin_unlock(&vmap_area_lock); + } nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr); @@ -1871,6 +1873,19 @@ struct vmap_area *find_vmap_area(unsigned long addr) return va; } +static struct vmap_area *find_unlink_vmap_area(unsigned long addr) +{ + struct vmap_area *va; + + spin_lock(&vmap_area_lock); + va = __find_vmap_area(addr, &vmap_area_root); + if (va) + unlink_va(va, &vmap_area_root); + spin_unlock(&vmap_area_lock); + + return va; +} + /*** Per cpu kva allocator ***/ /* @@ -2591,6 +2606,20 @@ struct vm_struct *find_vm_area(const void *addr) return va->vm; } +static struct vm_struct *__remove_vm_area(struct vmap_area *va) +{ + struct vm_struct *vm; + + if (!va || !va->vm) + return NULL; + + vm = va->vm; + kasan_free_module_shadow(vm); + free_unmap_vmap_area(va); + + return vm; +} + /** * remove_vm_area - find and remove a continuous kernel virtual area * @addr: base address @@ -2607,22 +2636,8 @@ struct vm_struct *remove_vm_area(const void *addr) might_sleep(); - spin_lock(&vmap_area_lock); - va = __find_vmap_area((unsigned long)addr, &vmap_area_root); - if (va && va->vm) { - struct vm_struct *vm = va->vm; - - va->vm = NULL; - spin_unlock(&vmap_area_lock); - - kasan_free_module_shadow(vm); - free_unmap_vmap_area(va); - - return vm; - } - - spin_unlock(&vmap_area_lock); - return NULL; + va = find_unlink_vmap_area((unsigned long) addr); + return __remove_vm_area(va); } static inline void set_area_direct_map(const struct vm_struct *area, @@ -2637,15 +2652,16 @@ static inline void set_area_direct_map(const struct vm_struct *area, } /* Handle removing and resetting vm mappings related to the vm_struct. */ -static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) +static void vm_remove_mappings(struct vmap_area *va, int deallocate_pages) { + struct vm_struct *area = va->vm; unsigned long start = ULONG_MAX, end = 0; unsigned int page_order = vm_area_page_order(area); int flush_reset = area->flags & VM_FLUSH_RESET_PERMS; int flush_dmap = 0; int i; - remove_vm_area(area->addr); + __remove_vm_area(va); /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */ if (!flush_reset) @@ -2690,6 +2706,7 @@ static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages) static void __vunmap(const void *addr, int deallocate_pages) { struct vm_struct *area; + struct vmap_area *va; if (!addr) return; @@ -2698,19 +2715,20 @@ static void __vunmap(const void *addr, int deallocate_pages) addr)) return; - area = find_vm_area(addr); - if (unlikely(!area)) { + va = find_unlink_vmap_area((unsigned long)addr); + if (unlikely(!va)) { WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n", addr); return; } + area = va->vm; debug_check_no_locks_freed(area->addr, get_vm_area_size(area)); debug_check_no_obj_freed(area->addr, get_vm_area_size(area)); kasan_poison_vmalloc(area->addr, get_vm_area_size(area)); - vm_remove_mappings(area, deallocate_pages); + vm_remove_mappings(va, deallocate_pages); if (deallocate_pages) { int i;
Currently __vunmap() path calls __find_vmap_area() two times. One on entry to check that area exists, second time inside remove_vm_area() function that also performs a new search of VA. In order to improvie it from a performance point of view we split remove_vm_area() into two new parts: - find_unlink_vmap_area() that does a search and unlink from tree; - __remove_vm_area() that does a removing but without searching. In this case there is no any functional change for remove_vm_area() whereas vm_remove_mappings(), where a second search happens, switches to the __remove_vm_area() variant where already detached VA is passed as a parameter, so there is no need to find it again. Performance wise, i use test_vmalloc.sh with 32 threads doing alloc free on a 64-CPUs-x86_64-box: perf without this patch: - 31.41% 0.50% vmalloc_test/10 [kernel.vmlinux] [k] __vunmap - 30.92% __vunmap - 17.67% _raw_spin_lock native_queued_spin_lock_slowpath - 12.33% remove_vm_area - 11.79% free_vmap_area_noflush - 11.18% _raw_spin_lock native_queued_spin_lock_slowpath 0.76% free_unref_page perf with this patch: - 11.35% 0.13% vmalloc_test/14 [kernel.vmlinux] [k] __vunmap - 11.23% __vunmap - 8.28% find_unlink_vmap_area - 7.95% _raw_spin_lock 7.44% native_queued_spin_lock_slowpath - 1.93% free_vmap_area_noflush - 0.56% _raw_spin_lock 0.53% native_queued_spin_lock_slowpath 0.60% __vunmap_range_noflush __vunmap() consumes around ~20% less CPU cycles on this test. Reported-by: Roman Gushchin <roman.gushchin@linux.dev> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> --- mm/vmalloc.c | 66 +++++++++++++++++++++++++++++++++------------------- 1 file changed, 42 insertions(+), 24 deletions(-)