path: root/src/lib/utils/mem_pool/mem_pool.cpp

/*
* (C) 2018,2019 Jack Lloyd
*
* Botan is released under the Simplified BSD License (see license.txt)
*/

#include <botan/internal/mem_pool.h>
#include <botan/mem_ops.h>
#include <algorithm>

#if defined(BOTAN_MEM_POOL_USE_MMU_PROTECTIONS)
   #include <botan/internal/os_utils.h>
#endif

namespace Botan {

/*
* Memory pool theory of operation
*
* This allocator is not useful for general purpose but works well within the
* context of allocating cryptographic keys. It makes several assumptions which
* don't work for implementing malloc but simplify and speed up the implementation:
*
* - There is some set of pages, which cannot be expanded later. These are pages
*   which were allocated, mlocked and passed to the Memory_Pool constructor.
*
* - The allocator is allowed to return null anytime it feels like not servicing
*   a request, in which case the request will be sent to calloc instead. In
*   particular, requests which are too small or too large are rejected.
*
* - Most allocations are powers of 2, the remainder are usually a multiple of 8
*
* - Free requests include the size of the allocation, so there is no need to
*   track this within the pool.
*
* - Alignment is important to the caller. For this allocator, any allocation of
*   size N is aligned evenly at N bytes.
*
* Initially each page is in the free page list. Each page is used for just one
* size of allocation, with requests bucketed into a small number of common
* sizes. If the allocation would be too big or too small it is rejected by the pool.
*
* The free list is maintained by a bitmap, one per page/Bucket. Since each
* Bucket only maintains objects of a single size, each bit set or clear
* indicates the status of one object.
*
* An allocation walks the list of buckets and asks each in turn if there is
* space. If a Bucket does not have any space, it sets a boolean flag m_is_full
* so that it does not need to rescan when asked again. The flag is cleared on
* first free from that bucket. If no bucket has space, but there are some free
* pages left, a free page is claimed as a new Bucket for that size. In this case
* it is pushed to the front of the list so it is first in line to service new
* requests.
*
* A deallocation also walks the list of buckets for the size and asks each
* Bucket in turn if it recognizes the pointer. When a Bucket becomes empty as a
* result of a deallocation, it is recycled back into the free pool. When this
* happens, the Buckets page goes to the end of the free list. All pages on the
* free list are marked in the MMU as noaccess, so anything touching them will
* immediately crash. They are only marked R/W once placed into a new bucket.
* Making the free list FIFO maximizes the time between the last free of a bucket
* and that page being writable again, maximizing chances of crashing after a
* use-after-free.
*
* Future work
* -------------
*
* The allocator is protected by a global lock. It would be good to break this
* up, since almost all of the work can actually be done in parallel especially
* when allocating objects of different sizes (which can't possibly share a
* bucket).
*
* It may be worthwhile to optimize deallocation by storing the Buckets in order
* (by pointer value) which would allow binary search to find the owning bucket.
*
* A useful addition would be to randomize the allocations. Memory_Pool would be
* changed to receive also a RandomNumberGenerator& object (presumably the system
* RNG, or maybe a ChaCha_RNG seeded with system RNG). Then the bucket to use and
* the offset within the bucket would be chosen randomly, instead of using first fit.
*
* Right now we don't make any provision for threading, so if two threads both
* allocate 32 byte values one after the other, the two allocations will likely
* share a cache line. Ensuring that distinct threads will (tend to) use distinct
* buckets would reduce this.
*
* Supporting a realloc-style API may be useful.
*/

namespace {

size_t choose_bucket(size_t n)
   {
   const size_t MINIMUM_ALLOCATION = 16;
   const size_t MAXIMUM_ALLOCATION = 256;

   if(n < MINIMUM_ALLOCATION || n > MAXIMUM_ALLOCATION)
      return 0;

   // Need to tune these

   const size_t buckets[] = {
      16, 24, 32, 48, 64, 80, 96, 112, 128, 160, 192, 256, 0,
   };

   for(size_t i = 0; buckets[i]; ++i)
      {
      if(n <= buckets[i])
         {
         return buckets[i];
         }
      }

   return 0;
   }

inline bool ptr_in_pool(const void* pool_ptr, size_t poolsize,
                        const void* buf_ptr, size_t bufsize)
   {
   const uintptr_t pool = reinterpret_cast<uintptr_t>(pool_ptr);
   const uintptr_t buf = reinterpret_cast<uintptr_t>(buf_ptr);
   return (buf >= pool) && (buf + bufsize <= pool + poolsize);
   }

// return index of first set bit
template<typename T>
size_t find_set_bit(T b)
   {
   size_t s = 8*sizeof(T) / 2;
   size_t bit = 0;

   // In this context we don't need to be const-time
   while(s > 0)
      {
      const T mask = (static_cast<T>(1) << s) - 1;
      if((b & mask) == 0)
         {
         bit += s;
         b >>= s;
         }
      s /= 2;
      }

   return bit;
   }

class BitMap final
   {
   public:
      BitMap(size_t bits) : m_len(bits)
         {
         m_bits.resize((bits + BITMASK_BITS - 1) / BITMASK_BITS);
         m_main_mask = static_cast<bitmask_type>(~0);
         m_last_mask = m_main_mask;

         if(bits % BITMASK_BITS != 0)
            m_last_mask = (static_cast<bitmask_type>(1) << (bits % BITMASK_BITS)) - 1;
         }

      bool find_free(size_t* bit);

      void free(size_t bit)
         {
         BOTAN_ASSERT_NOMSG(bit <= m_len);
         const size_t w = bit / BITMASK_BITS;
         BOTAN_ASSERT_NOMSG(w < m_bits.size());
         const bitmask_type mask = static_cast<bitmask_type>(1) << (bit % BITMASK_BITS);
         m_bits[w] = m_bits[w] & (~mask);
         }

      bool empty() const
         {
         for(auto bitset : m_bits)
            {
            if(bitset != 0)
               {
               return false;
               }
            }

         return true;
         }

  private:
#if defined(BOTAN_ENABLE_DEBUG_ASSERTS)
      typedef uint8_t bitmask_type;
      enum { BITMASK_BITS = 8 };
#else
      typedef word bitmask_type;
      enum { BITMASK_BITS = BOTAN_MP_WORD_BITS };
#endif

      size_t m_len;
      bitmask_type m_main_mask;
      bitmask_type m_last_mask;
      std::vector<bitmask_type> m_bits;
   };

bool BitMap::find_free(size_t* bit)
   {
   for(size_t i = 0; i != m_bits.size(); ++i)
      {
      const bitmask_type mask = (i == m_bits.size() - 1) ? m_last_mask : m_main_mask;
      if((m_bits[i] & mask) != mask)
         {
         const size_t free_bit = find_set_bit(~m_bits[i]);
         const bitmask_type bmask = static_cast<bitmask_type>(1) << (free_bit % BITMASK_BITS);
         BOTAN_ASSERT_NOMSG((m_bits[i] & bmask) == 0);
         m_bits[i] |= bmask;
         *bit = BITMASK_BITS*i + free_bit;
         return true;
         }
      }

   return false;
   }

}

class Bucket final
   {
   public:
      Bucket(uint8_t* mem, size_t mem_size, size_t item_size) :
         m_item_size(item_size),
         m_page_size(mem_size),
         m_range(mem),
         m_bitmap(mem_size / item_size),
         m_is_full(false)
         {
         }

      uint8_t* alloc()
         {
         if(m_is_full)
            {
            // I know I am full
            return nullptr;
            }

         size_t offset;
         if(!m_bitmap.find_free(&offset))
            {
            // I just found out I am full
            m_is_full = true;
            return nullptr;
            }

         BOTAN_ASSERT(offset * m_item_size < m_page_size, "Offset is in range");
         return m_range + m_item_size*offset;
         }

      bool free(void* p)
         {
         if(!in_this_bucket(p))
            return false;

         /*
         Zero also any trailing bytes, which should not have been written to,
         but maybe the user was bad and wrote past the end.
         */
         std::memset(p, 0, m_item_size);

         const size_t offset = (reinterpret_cast<uintptr_t>(p) - reinterpret_cast<uintptr_t>(m_range)) / m_item_size;

         m_bitmap.free(offset);
         m_is_full = false;

         return true;
         }

      bool in_this_bucket(void* p) const
         {
         return ptr_in_pool(m_range, m_page_size, p, m_item_size);
         }

      bool empty() const
         {
         return m_bitmap.empty();
         }

      uint8_t* ptr() const
         {
         return m_range;
         }

   private:
      size_t m_item_size;
      size_t m_page_size;
      uint8_t* m_range;
      BitMap m_bitmap;
      bool m_is_full;
   };

Memory_Pool::Memory_Pool(const std::vector<void*>& pages, size_t page_size) :
   m_page_size(page_size)
   {
   m_min_page_ptr = ~static_cast<uintptr_t>(0);
   m_max_page_ptr = 0;

   for(auto page : pages)
      {
      const uintptr_t p = reinterpret_cast<uintptr_t>(page);

      m_min_page_ptr = std::min(p, m_min_page_ptr);
      m_max_page_ptr = std::max(p, m_max_page_ptr);

      clear_bytes(page, m_page_size);
#if defined(BOTAN_MEM_POOL_USE_MMU_PROTECTIONS)
      OS::page_prohibit_access(pages[i]);
#endif
      m_free_pages.push_back(static_cast<uint8_t*>(page));
      }

   /*
   Right now this points to the start of the last page, adjust it to instead
   point to the first byte of the following page
   */
   m_max_page_ptr += page_size;
   }

Memory_Pool::~Memory_Pool()
   {
#if defined(BOTAN_MEM_POOL_USE_MMU_PROTECTIONS)
   for(size_t i = 0; i != m_free_pages.size(); ++i)
      {
      OS::page_allow_access(m_free_pages[i]);
      }
#endif
   }

void* Memory_Pool::allocate(size_t n)
   {
   if(n > m_page_size)
      return nullptr;

   const size_t n_bucket = choose_bucket(n);

   if(n_bucket > 0)
      {
      lock_guard_type<mutex_type> lock(m_mutex);

      std::deque<Bucket>& buckets = m_buckets_for[n_bucket];

      /*
      It would be optimal to pick the bucket with the most usage,
      since a bucket with say 1 item allocated out of it has a high
      chance of becoming later freed and then the whole page can be
      recycled.
      */
      for(auto& bucket : buckets)
         {
         if(uint8_t* p = bucket.alloc())
            return p;

         // If the bucket is full, maybe move it to the end of the list?
         // Otoh bucket search should be very fast
         }

      if(!m_free_pages.empty())
         {
         uint8_t* ptr = m_free_pages[0];
         m_free_pages.pop_front();
#if defined(BOTAN_MEM_POOL_USE_MMU_PROTECTIONS)
         OS::page_allow_access(ptr);
#endif
         buckets.push_front(Bucket(ptr, m_page_size, n_bucket));
         void* p = buckets[0].alloc();
         BOTAN_ASSERT_NOMSG(p != nullptr);
         return p;
         }
      }

   // out of room
   return nullptr;
   }

bool Memory_Pool::deallocate(void* p, size_t len) noexcept
   {
   // Do a fast range check first, before taking the lock
   const uintptr_t p_val = reinterpret_cast<uintptr_t>(p);
   if(p_val < m_min_page_ptr || p_val > m_max_page_ptr)
      return false;

   const size_t n_bucket = choose_bucket(len);

   if(n_bucket != 0)
      {
      try
         {
         lock_guard_type<mutex_type> lock(m_mutex);

         std::deque<Bucket>& buckets = m_buckets_for[n_bucket];

         for(size_t i = 0; i != buckets.size(); ++i)
            {
            Bucket& bucket = buckets[i];
            if(bucket.free(p))
               {
               if(bucket.empty())
                  {
#if defined(BOTAN_MEM_POOL_USE_MMU_PROTECTIONS)
                  OS::page_prohibit_access(bucket.ptr());
#endif
                  m_free_pages.push_back(bucket.ptr());

                  if(i != buckets.size() - 1)
                     std::swap(buckets.back(), buckets[i]);
                  buckets.pop_back();
                  }
               return true;
               }
            }
         }
      catch(...)
         {
         /*
         * The only exception throws that can occur in the above code are from
         * either the STL or BOTAN_ASSERT failures. In either case, such an
         * error indicates a logic error or data corruption in the memory
         * allocator such that it is no longer safe to continue executing.
         *
         * Since this function is noexcept, simply letting the exception escape
         * is sufficient for terminate to be called. However in this scenario
         * it is implementation defined if any stack unwinding is performed.
         * Since stack unwinding could cause further memory deallocations this
         * could result in further corruption in this allocator state. To prevent
         * this, call terminate directly.
         */
         std::terminate();
         }
      }

   return false;
   }

}