On Wed, Dec 07, 2022 at 04:03:41PM +0800, Baoquan He wrote:
> On 12/05/22 at 01:56pm, Uladzislau Rezki wrote:
> > > Through vmalloc API, a virtual kernel area is reserved for physical
> > > address mapping. And vmap_area is used to track them, while vm_struct
> > > is allocated to associate with the vmap_area to store more information
> > > and passed out.
> > >
> > > However, area reserved via vm_map_ram() is an exception. It doesn't have
> > > vm_struct to associate with vmap_area. And we can't recognize the
> > > vmap_area with '->vm == NULL' as a vm_map_ram() area because the normal
> > > freeing path will set va->vm = NULL before unmapping, please see
> > > function remove_vm_area().
> > >
> > > Meanwhile, there are two types of vm_map_ram area. One is the whole
> > > vmap_area being reserved and mapped at one time; the other is the
> > > whole vmap_area with VMAP_BLOCK_SIZE size being reserved, while mapped
> > > into split regions with smaller size several times via vb_alloc().
> > >
> > > To mark the area reserved through vm_map_ram(), add flags field into
> > > struct vmap_area. Bit 0 indicates whether it's a vm_map_ram area,
> > > while bit 1 indicates whether it's a vmap_block type of vm_map_ram
> > > area.
> > >
> > > This is a preparatoin for later use.
> > >
> > > Signed-off-by: Baoquan He <bhe@xxxxxxxxxx>
> > > ---
> > > include/linux/vmalloc.h | 1 +
> > > mm/vmalloc.c | 18 +++++++++++++++++-
> > > 2 files changed, 18 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
> > > index 096d48aa3437..69250efa03d1 100644
> > > --- a/include/linux/vmalloc.h
> > > +++ b/include/linux/vmalloc.h
> > > @@ -76,6 +76,7 @@ struct vmap_area {
> > > unsigned long subtree_max_size; /* in "free" tree */
> > > struct vm_struct *vm; /* in "busy" tree */
> > > };
> > > + unsigned long flags; /* mark type of vm_map_ram area */
> > > };
> > >
> > > /* archs that select HAVE_ARCH_HUGE_VMAP should override one or more of these */
> > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > index 5d3fd3e6fe09..d6f376060d83 100644
> > > --- a/mm/vmalloc.c
> > > +++ b/mm/vmalloc.c
> > > @@ -1815,6 +1815,7 @@ static void free_vmap_area_noflush(struct vmap_area *va)
> > >
> > > spin_lock(&vmap_area_lock);
> > > unlink_va(va, &vmap_area_root);
> > > + va->flags = 0;
> > > spin_unlock(&vmap_area_lock);
> > >
> > This is not a good place to set flags to zero. It looks to me like
> > corner and kind of specific.
>
> Thanks for reviewing.
>
> Here, I thought to clear VMAP_RAM|VMAP_BLOCK on vmap->flags when free
> the vmap_block. I didn't find a good place to do the clearing. When we
> call free_vmap_block(), we either come from purge_fragmented_blocks(),
> or from vb_free(). In vb_free(), it will call free_vmap_block() when
> the whole vmap_block is dirty. In purge_fragmented_blocks(), it will
> try to purge all vmap_block which only has dirty or free regions.
> For both of above functions, they will call free_vmap_block() when
> there's no being used region in the vmap_block.
>
> purge_fragmented_blocks()
> vb_free()
> -->free_vmap_block()
>
> So seems we don't need to clear the VMAP_RAM|VMAP_BLOCK on vmap->flags
> because there's no mapping existed in the vmap_block. The consequent
> free_vmap_block() will remove the relevant vmap_area from vmap_area_list
> and vmap_area_root tree.
>
> So I plan to remove code change in this place.
> >
> >
> > > nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
> > > @@ -1887,6 +1888,10 @@ struct vmap_area *find_vmap_area(unsigned long addr)
> > >
> > > #define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
> > >
> > > +#define VMAP_RAM 0x1
> > > +#define VMAP_BLOCK 0x2
> > > +#define VMAP_FLAGS_MASK 0x3
> > > +
> > > struct vmap_block_queue {
> > > spinlock_t lock;
> > > struct list_head free;
> > > @@ -1967,6 +1972,9 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
> > > kfree(vb);
> > > return ERR_CAST(va);
> > > }
> > > + spin_lock(&vmap_area_lock);
> > > + va->flags = VMAP_RAM|VMAP_BLOCK;
> > > + spin_unlock(&vmap_area_lock);
> > >
> > The per-cpu code was created as a fast per-cpu allocator because of high
> > vmalloc lock contention. If possible we should avoid of locking of the
> > vmap_area_lock. Because it has a high contention.
>
> Fair enough. I made below draft patch to address the concern. By
> adding argument va_flags to alloc_vmap_area(), we can pass the
> vm_map_ram flags into alloc_vmap_area and filled into vmap_area->flags.
> With this, we don't need add extra action to acquire vmap_area_root lock
> and do the flags setting. Is it OK to you?
>
> From 115f6080b339d0cf9dd20c5f6c0d3121f6b22274 Mon Sep 17 00:00:00 2001
> From: Baoquan He <bhe@xxxxxxxxxx>
> Date: Wed, 7 Dec 2022 11:08:14 +0800
> Subject: [PATCH] mm/vmalloc: change alloc_vmap_area() to pass in va_flags
>
> With this change, we can pass and set vmap_area->flags for vm_map_ram area
> in alloc_vmap_area(). Then no extra action need be added to acquire
> vmap_area_lock when doing the vmap_area->flags setting.
>
> Signed-off-by: Baoquan He <bhe@xxxxxxxxxx>
> ---
> mm/vmalloc.c | 13 +++++++++----
> 1 file changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index ccaa461998f3..d74eddec352f 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -1586,7 +1586,9 @@ preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
> static struct vmap_area *alloc_vmap_area(unsigned long size,
> unsigned long align,
> unsigned long vstart, unsigned long vend,
> - int node, gfp_t gfp_mask)
> + int node, gfp_t gfp_mask,
> + unsigned long va_flags)
> +)
> {
> struct vmap_area *va;
> unsigned long freed;
> @@ -1630,6 +1632,7 @@ static struct vmap_area *alloc_vmap_area(unsigned long size,
> va->va_start = addr;
> va->va_end = addr + size;
> va->vm = NULL;
> + va->flags = va_flags;
>
> spin_lock(&vmap_area_lock);
> insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
> @@ -1961,7 +1964,8 @@ static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
>
> va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
> VMALLOC_START, VMALLOC_END,
> - node, gfp_mask);
> + node, gfp_mask,
> + VMAP_RAM|VMAP_BLOCK);
> if (IS_ERR(va)) {
> kfree(vb);
> return ERR_CAST(va);
> @@ -2258,7 +2262,8 @@ void *vm_map_ram(struct page **pages, unsigned int count, int node)
> } else {
> struct vmap_area *va;
> va = alloc_vmap_area(size, PAGE_SIZE,
> - VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
> + VMALLOC_START, VMALLOC_END,
> + node, GFP_KERNEL, VMAP_RAM|VMAP_BLOCK);
> if (IS_ERR(va))
> return NULL;
>
> @@ -2498,7 +2503,7 @@ static struct vm_struct *__get_vm_area_node(unsigned long size,
> if (!(flags & VM_NO_GUARD))
> size += PAGE_SIZE;
>
> - va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
> + va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0);
> if (IS_ERR(va)) {
> kfree(area);
> return NULL;
> --
> 2.34.1
>
Yes, this is better than it was before. Adding an extra parameter makes
it more valid and logical.
--
Uladzislau Rezki
