[PATCH 1/1] lib: sbi_scratch: re-implement scratch memory allocator

Anup Patel anup at brainfault.org
Tue Jun 22 05:01:24 PDT 2021


Hi Heinrich,

On Wed, Jun 9, 2021 at 2:01 AM Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>
> Up to now we could allocated scratch memory but not deallocate it.
>
> Provide a best fit memory allocator.
>
> Signed-off-by: Heinrich Schuchardt <xypron.glpk at gmx.de>
> ---
>  include/sbi/sbi_scratch.h |  52 +++++++++-
>  include/sbi/sbi_types.h   |   2 +
>  lib/sbi/sbi_scratch.c     | 203 +++++++++++++++++++++++++++++++-------
>  3 files changed, 218 insertions(+), 39 deletions(-)
>
> diff --git a/include/sbi/sbi_scratch.h b/include/sbi/sbi_scratch.h
> index 186a40c..f9ea434 100644
> --- a/include/sbi/sbi_scratch.h
> +++ b/include/sbi/sbi_scratch.h
> @@ -36,10 +36,10 @@
>  #define SBI_SCRATCH_TMP0_OFFSET                        (9 * __SIZEOF_POINTER__)
>  /** Offset of options member in sbi_scratch */
>  #define SBI_SCRATCH_OPTIONS_OFFSET             (10 * __SIZEOF_POINTER__)
> -/** Offset of extra space in sbi_scratch */
> -#define SBI_SCRATCH_EXTRA_SPACE_OFFSET         (11 * __SIZEOF_POINTER__)
>  /** Maximum size of sbi_scratch (4KB) */
>  #define SBI_SCRATCH_SIZE                       (0x1000)
> +/** Offset of field mem in the sbi_mem_alloc structure */
> +#define SBI_SCRATCH_ALLOC_SIZE (offsetof(struct sbi_scratch_alloc, mem))

The SBI_SCRATCH_ALLOC_SIZE is not assembly friendly so should be under
"#ifndef __ASSEMBLER__". Better to place it just after "struct
sbi_scratch_alloc"
definition.

>
>  /* clang-format on */
>
> @@ -47,6 +47,49 @@
>
>  #include <sbi/sbi_types.h>
>
> +/**
> + * struct sbi_scratch_alloc - memory allocation
> + *
> + * This structure describes a block of allocated or free memory.
> + * The fields @prev and @next are only used for free blocks.
> + */
> +struct sbi_scratch_alloc {
> +       /**
> +        * @prev_size: size of previous memory block
> +        *
> +        * If the bit 0 is zero the memory is available.
> +        * If the bit 0 is non-zero the memory is allocated.
> +        */
> +       unsigned int prev_size;
> +       /**
> +        * @size: size of memory block
> +        *
> +        * If the bit 0 is zero the memory is available.
> +        * If the bit 0 is non-zero the memory is allocated.
> +        */
> +       unsigned int size;
> +
> +       union {
> +               /**
> +                * @mem: allocated memory
> +                *
> +                * The macro SBI_SCRATCH_ALLOC_SIZE provides the offset of @mem
> +                * in the sbi_mem_alloc structure.
> +                */
> +               unsigned char mem[0];
> +               struct {
> +                       /**
> +                        * @prev: offset of preceeding block in free block list
> +                        */
> +                       unsigned int prev;
> +                       /**
> +                        * @next: offset of succceeding block in free block list
> +                        */
> +                       unsigned int next;
> +               };
> +       };
> +} __aligned(2 * sizeof(int));

I would suggest to have
#define SBI_SCRATCH_ALLOC_ALIGN (2 * sizeof(unsigned int))
and use it in above __aligned() and other places in sbi_scratch.c

> +
>  /** Representation of per-HART scratch space */
>  struct sbi_scratch {
>         /** Start (or base) address of firmware linked to OpenSBI library */
> @@ -71,6 +114,8 @@ struct sbi_scratch {
>         unsigned long tmp0;
>         /** Options for OpenSBI library */
>         unsigned long options;
> +       /** Start of scratch memory allocation list */
> +       struct sbi_scratch_alloc mem;
>  };
>
>  /** Possible options for OpenSBI library */
> @@ -95,8 +140,7 @@ int sbi_scratch_init(struct sbi_scratch *scratch);
>  /**
>   * Allocate from extra space in sbi_scratch
>   *
> - * @return zero on failure and non-zero (>= SBI_SCRATCH_EXTRA_SPACE_OFFSET)
> - * on success
> + * @return zero on failure and non-zero on success
>   */
>  unsigned long sbi_scratch_alloc_offset(unsigned long size);
>
> diff --git a/include/sbi/sbi_types.h b/include/sbi/sbi_types.h
> index 38e3565..2050265 100644
> --- a/include/sbi/sbi_types.h
> +++ b/include/sbi/sbi_types.h
> @@ -88,6 +88,8 @@ typedef unsigned long         physical_size_t;
>  #define STR(x) XSTR(x)
>  #define XSTR(x) #x
>
> +#define ALIGN(x, a) ((typeof(x))((unsigned long)(x + (a - 1)) & ~(a - 1)))
> +
>  #define ROUNDUP(a, b) ((((a)-1) / (b) + 1) * (b))
>  #define ROUNDDOWN(a, b) ((a) / (b) * (b))
>
> diff --git a/lib/sbi/sbi_scratch.c b/lib/sbi/sbi_scratch.c
> index 87b34c6..4c59c62 100644
> --- a/lib/sbi/sbi_scratch.c
> +++ b/lib/sbi/sbi_scratch.c
> @@ -18,10 +18,16 @@ u32 last_hartid_having_scratch = SBI_HARTMASK_MAX_BITS - 1;
>  struct sbi_scratch *hartid_to_scratch_table[SBI_HARTMASK_MAX_BITS] = { 0 };
>
>  static spinlock_t extra_lock = SPIN_LOCK_INITIALIZER;
> -static unsigned long extra_offset = SBI_SCRATCH_EXTRA_SPACE_OFFSET;
> +static unsigned int first_free;
>
>  typedef struct sbi_scratch *(*hartid2scratch)(ulong hartid, ulong hartindex);
>
> +/**
> + * sbi_scratch_init() - initialize scratch table and allocator
> + *
> + * @scratch:   pointer to table
> + * Return:     0 on success
> + */
>  int sbi_scratch_init(struct sbi_scratch *scratch)
>  {
>         u32 i;
> @@ -37,63 +43,190 @@ int sbi_scratch_init(struct sbi_scratch *scratch)
>                         last_hartid_having_scratch = i;
>         }
>
> +       /* Initialize memory allocation block list */
> +       scratch = sbi_hartid_to_scratch(last_hartid_having_scratch);
> +
> +       scratch->mem.prev_size = (2 * sizeof(unsigned int)) | 1U;
> +       scratch->mem.size = SBI_SCRATCH_SIZE -
> +                           offsetof(struct sbi_scratch, mem.mem);
> +       first_free = offsetof(struct sbi_scratch, mem);
> +       scratch->mem.prev = 0;
> +       scratch->mem.next = 0;
> +
>         return 0;
>  }
>
> +/**
> + * sbi_scratch_alloc_offset() - allocate scratch memory
> + *
> + * @size:      requested size
> + * Return:     offset of allocated block on succcess, 0 on failure
> + */
>  unsigned long sbi_scratch_alloc_offset(unsigned long size)
>  {
> -       u32 i;
> -       void *ptr;
> -       unsigned long ret = 0;
> -       struct sbi_scratch *rscratch;
> +       unsigned long ret;
> +       unsigned int best_size = ~0U;
> +       struct sbi_scratch_alloc *best = NULL;
> +       struct sbi_scratch *scratch =
> +               sbi_hartid_to_scratch(last_hartid_having_scratch);
> +       unsigned int next;
> +       struct sbi_scratch_alloc *current;
> +       struct sbi_scratch_alloc *pred, *succ;
> +       struct sbi_scratch_alloc *end =
> +               (void *)((char *)scratch + SBI_SCRATCH_SIZE);
>
>         /*
> -        * We have a simple brain-dead allocator which never expects
> -        * anything to be free-ed hence it keeps incrementing the
> -        * next allocation offset until it runs-out of space.
> -        *
> -        * In future, we will have more sophisticated allocator which
> -        * will allow us to re-claim free-ed space.
> +        * When allocating zero bytes we still need space
> +        * for the prev and next fields.
>          */
> -
>         if (!size)
> +               size = 1;
> +       size = ALIGN(size, 2 * sizeof(unsigned int));
> +
> +       spin_lock(&extra_lock);
> +
> +       /* Find best fitting free block */
> +       for (next = first_free; next; next = current->next) {
> +               current = (void *)((char *)scratch + next);
> +               if (current->size > best_size || current->size < size)
> +                       continue;
> +               best_size = current->size;
> +               best = current;
> +       }
> +       if (!best) {
> +               spin_unlock(&extra_lock);
>                 return 0;
> +       }
>
> -       if (size & (__SIZEOF_POINTER__ - 1))
> -               size = (size & ~(__SIZEOF_POINTER__ - 1)) + __SIZEOF_POINTER__;
> +       /* Update free list */
> +       if (best->prev)
> +               pred = (void *)((char *)scratch + best->prev);
> +       else
> +               pred = NULL;
> +       if (best->next)
> +               succ = (void *)((char *)scratch + best->next);
> +       else
> +               succ = NULL;
> +
> +       if (best->size > size + SBI_SCRATCH_ALLOC_SIZE) {

We should split a block only if we have enough residue memory for
prev_size, size, prev, and next members of struct sbi_scratch_alloc.

I think this if-condition should be:
"if (best->size >= (size + 2 * SBI_SCRATCH_ALLOC_SIZE))"

> +               /* Split block, use the lower part for allocation. */
> +               current = (struct sbi_scratch_alloc *)&best->mem[size];
> +               next = (char *)current - (char *)scratch;
> +               current->size = best->size - size -
> +                               SBI_SCRATCH_ALLOC_SIZE;
> +               current->prev = best->prev;
> +               current->next = best->next;
> +               current->prev_size = size | 1U;

Please add a #define for this "1U".

> +               best->size = size;
> +               if (succ)
> +                       succ->prev = next;
> +       } else {
> +               next = best->next;
> +               if (succ)
> +                       succ->prev = best->prev;
> +               current = best;
> +       }
>
> -       spin_lock(&extra_lock);
> +       if (pred)
> +               pred->next = next;
> +       else
> +               first_free = next;
>
> -       if (SBI_SCRATCH_SIZE < (extra_offset + size))
> -               goto done;
> +       /* Update memory block list */
> +       succ = (struct sbi_scratch_alloc *)&current->mem[current->size];
>
> -       ret = extra_offset;
> -       extra_offset += size;
> +       best->size |= 1U;

Same as above, need a #define for "1U" here as well.

>
> -done:
> -       spin_unlock(&extra_lock);
> +       if (succ < end)
> +               succ->prev_size = current->size;
>
> -       if (ret) {
> -               for (i = 0; i <= sbi_scratch_last_hartid(); i++) {
> -                       rscratch = sbi_hartid_to_scratch(i);
> -                       if (!rscratch)
> -                               continue;
> -                       ptr = sbi_scratch_offset_ptr(rscratch, ret);
> -                       sbi_memset(ptr, 0, size);
> -               }
> +       ret =  best->mem - (unsigned char *)scratch;
> +
> +       /* Erase allocated scratch memory */
> +       for (unsigned int i = 0; i <= last_hartid_having_scratch; i++) {
> +               void *ptr;
> +               struct sbi_scratch *rscratch;
> +
> +               rscratch = sbi_hartid_to_scratch(i);
> +               if (!rscratch)
> +                       continue;
> +               ptr = sbi_scratch_offset_ptr(rscratch, ret);
> +               sbi_memset(ptr, 0, size);

Any reason why we should zero-out memory with extra_lock held ??

>         }
>
> +       spin_unlock(&extra_lock);
> +
>         return ret;
>  }
>
> +/**
> + * sbi_scratch_free_offset() - free scratch memory
> + *
> + * @offset:    offset to memory to be freed
> + */
>  void sbi_scratch_free_offset(unsigned long offset)
>  {
> -       if ((offset < SBI_SCRATCH_EXTRA_SPACE_OFFSET) ||
> -           (SBI_SCRATCH_SIZE <= offset))
> +       struct sbi_scratch *scratch =
> +               sbi_hartid_to_scratch(last_hartid_having_scratch);
> +       struct sbi_scratch_alloc *freed = (void *)((unsigned char *)scratch +
> +                                     offset - SBI_SCRATCH_ALLOC_SIZE);
> +       struct sbi_scratch_alloc *pred, *succ;
> +       struct sbi_scratch_alloc *end =
> +               (void *)((char *)scratch + SBI_SCRATCH_SIZE);
> +
> +       spin_lock(&extra_lock);
> +
> +       if (!offset || !(freed->size & 1U)) {
> +               spin_unlock(&extra_lock);
>                 return;
> +       }
>
> -       /*
> -        * We don't actually free-up because it's a simple
> -        * brain-dead allocator.
> -        */
> +       /* Mark block as free */
> +       freed->size &= ~1U;
> +

Add single line comment here that we attempt to merge free block into
contiguous predecessor block.

> +       pred = (struct sbi_scratch_alloc *)((char *)freed -
> +              (freed->prev_size & ~1U) - SBI_SCRATCH_ALLOC_SIZE);
> +       if (pred >= &scratch->mem && !(pred->size & 1U)) {
> +               /* Coalesce free blocks */
> +               pred->size += freed->size + SBI_SCRATCH_ALLOC_SIZE;
> +               freed = pred;
> +       } else {
> +               /* Insert at start of free list */
> +               if (first_free) {
> +                       succ = (void *)((char *)scratch + first_free);
> +                       succ->prev = offset - SBI_SCRATCH_ALLOC_SIZE;
> +               }
> +               freed->next = first_free;
> +               freed->prev = 0;
> +               first_free = offset - SBI_SCRATCH_ALLOC_SIZE;
> +       }
> +

Add single line comment here that we attempt to merge contiguous
successor block into free block.

> +       succ = (struct sbi_scratch_alloc *)&freed->mem[freed->size & ~1U];
> +       if (succ < end) {
> +               if (!(succ->size & 1U)) {
> +                       struct sbi_scratch_alloc *succ2;
> +
> +                       /* Coalesce free blocks */
> +                       succ2 = (struct sbi_scratch_alloc *)
> +                               &succ->mem[succ->size & ~1U];
> +                       freed->size += SBI_SCRATCH_ALLOC_SIZE + succ->size;
> +                       if (succ2 < end)
> +                               succ2->prev_size = freed->size;
> +
> +                       /* Remove successor from free list */
> +                       if (succ->prev) {
> +                               pred = (void *)((char *)scratch + succ->prev);
> +                               pred->next = succ->next;
> +                       } else {
> +                               first_free = succ->next;
> +                       }
> +                       if (succ->next) {
> +                               succ2 = (void *)((char *)scratch + succ->next);
> +                               succ2->prev = succ->prev;
> +                       }
> +               } else {
> +                       succ->prev_size = freed->size;
> +               }
> +       }

Add a empty new line here.

> +       spin_unlock(&extra_lock);
>  }
> --
> 2.30.2
>
>
> --
> opensbi mailing list
> opensbi at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/opensbi

Apart from above minor comments, this patch looks good to me.

Great work !!!

Regards,
Anup



More information about the opensbi mailing list