summaryrefslogtreecommitdiff
path: root/PGU/CHAP9/memalloc.s
diff options
context:
space:
mode:
authorCarlos Maiolino <[email protected]>2025-07-10 22:55:07 +0200
committerCarlos Maiolino <[email protected]>2025-07-10 22:56:55 +0200
commitd98f46ce647846b0aa30b2e16a30fd4e152a1bf5 (patch)
tree267474fcc77cf20b428f6f4c7f768ca09f4cfe0e /PGU/CHAP9/memalloc.s
parent869e68986aa8f69af6e7842260a68d1e5c6f796f (diff)
Add new code
Signed-off-by: Carlos Maiolino <[email protected]>
Diffstat (limited to 'PGU/CHAP9/memalloc.s')
-rw-r--r--PGU/CHAP9/memalloc.s303
1 files changed, 303 insertions, 0 deletions
diff --git a/PGU/CHAP9/memalloc.s b/PGU/CHAP9/memalloc.s
new file mode 100644
index 0000000..6f354de
--- /dev/null
+++ b/PGU/CHAP9/memalloc.s
@@ -0,0 +1,303 @@
+
+# Alloc and Dealloc memory as requested
+#
+# Programs using these routines will ask
+# for a certain size of memory. We actually
+# use more than that size for metadata,
+# but we put it at the beginning, before the
+# pointer we hand back. We add a size field and
+# and AVAILABLE/UNAVAILABLE marker.
+#
+# The whole memory slot looks like this:
+#
+####################################################
+# Available marker | Size | Actual memory location #
+####################################################
+# ^-- returned pointer points
+# here
+#
+# The calling program won't see our metadata.
+
+### GLOBAL VARS ###
+.section .data
+
+############################### DEBUG INFO ################################
+
+# Originally, these labels heap_begin and current_break are .long long...
+# This implicitly means 4 bytes for each variable. This was causing the instructions
+# in allocate_init to overlap the variables when writing to them:
+
+# The following instructions were the problem:
+# movq %rax, current_break
+# allocated at 0x402011
+# movq %rax, heap_begin
+# allocated at 0x40200d
+#
+# Static allocation grows up in memory... But, when writing %rax to heap_begin,
+# due the variable sizes being 4bytes, the instruction was corrupting
+# current_break variable, causing the program to crash later in the allocation
+# loop, because the heap_begin allocation, consequently caused current_break to be
+# zero causing a NULL ptr dereference (or simply a SEGFAULT):
+
+# => 0x000000000040114e <+5>: mov 0x8(%rax),%rdx
+# Program received signal SIGSEGV, Segmentation fault.
+
+# allocate_init() disassemble:
+# 0x0000000000401105 <+0>: push %rbp
+# 0x0000000000401106 <+1>: mov %rsp,%rbp
+# 0x0000000000401109 <+4>: mov $0xc,%rax
+# 0x0000000000401110 <+11>: mov $0x0,%rdi
+# 0x0000000000401117 <+18>: syscall
+# 0x0000000000401119 <+20>: inc %rax
+# 0x000000000040111c <+23>: mov %rax,0x402011
+# 0x0000000000401124 <+31>: mov %rax,0x40200d
+#=> 0x000000000040112c <+39>: mov %rbp,%rsp
+# 0x000000000040112f <+42>: pop %rbp
+# 0x0000000000401130 <+43>: ret
+#
+# (gdb) info register rax
+# rax 0x403001 4206593
+# (gdb) p /x *0x40200d
+# $5 = 0x403001
+# (gdb) p /x *0x402011
+# $6 = 0x0 <-- Here, the address of current_break got zeroed after
+# heap_being has been changed
+
+############################# END OF DEBUG INFO #############################
+
+# Points to the beginning of the memory we are managing
+heap_begin:
+ .quad 0
+
+# Points to one locaiton past the memory we are managing
+current_break:
+ .quad 0
+
+### HEADER STRUCTURE INFORMATION ###
+
+# To make things simpler, we use one word for
+# each field in the header
+
+.equ HEADER_SIZE, 16 # size of space for memory region header
+.equ HDR_AVAIL_OFFSET, 0 # Location of the 'available' flag in the header
+.equ HDR_SIZE_OFFSET, 8 # Location of the size field in the header
+
+
+### CONSTANTS ###
+.equ UNAVAILABLE, 0
+.equ AVAILABLE, 1
+
+.equ SYS_BRK, 12 # syscall number for brk() syscall in x86_64
+
+
+.section .text
+
+### FUNCTIONS ###
+
+## allocate_init ##
+#
+# Call this function to initialize the functions by setting heap_begin and
+# current_break. No parameters and no return value
+
+.globl allocate_init
+.type allocate_init, @function
+
+allocate_init:
+ pushq %rbp # standard function stuff
+ movq %rsp, %rbp
+
+ # If brk() syscall is called with a 0 in %rdi, it returns the last valid
+ # usable address
+ movq $SYS_BRK, %rax
+ movq $0, %rdi
+ syscall
+
+ incq %rax # brk(0) returns the current break in %rax, we want the
+ # value after that
+
+ movq %rax, current_break # Store the current break (actually the address
+ # after that)
+
+ movq %rax, heap_begin # Our heap starts where the break is now. First
+ # address of uninitialized memory. This will
+ # cause the allocate function to get more memory
+ # from Linux the first time it is run.
+
+ movq %rbp, %rsp # Exit the function
+ popq %rbp
+ ret
+
+## END of allocate_init ##
+
+
+## allocate ##
+#
+# Grab a section of memory.
+# - Checks to see if there are any free blocks
+# - If not, asks Linux for more memory for the
+# heap through brk() syscall
+#
+# - Receives on parameter, the size of the memory block
+# we want to allocate
+#
+# - Returns the address of the allocated memory in %rax, or 0
+# if there is no memory available on the system
+#
+# ### Process
+#
+# %rcx - holds the size of requested memory (first/only parameter)
+# %rax - current memory region being examined
+# %rbx - Memory address past the end of the heap
+# %rdx - size of the current memory region
+#
+# We scan through each memory region starting with heap_begin. We look at the
+# size of each one and if it has been allocated. If it's big enough for the
+# requested size, and it's available, we grab that one. If we do not find a
+# region large enough, we ask Linux for more memory, which in case, moves the
+# current_break up.
+
+.globl allocate
+.type allocate, @function
+.equ ST_MEM_SIZE, 16 # Stack position of the memory size to allocate
+
+allocate:
+ pushq %rbp
+ movq %rsp, %rbp
+
+ movq ST_MEM_SIZE(%rbp), %rcx # %rcx now holds the memory size we are
+ # looking for (first/only parameter)
+
+ movq heap_begin, %rax # %rax will hold the current search location
+ movq current_break, %rbx
+
+ alloc_loop_begin: # Scan memory regions
+
+ cmpq %rbx, %rax # heap needs more memory if these are
+ # equal. %rax will hold the end of the
+ # next memory region at each iteraction
+ # until it hits the current_break.
+
+ je move_break
+
+ # Retrieve the size of this mem slot
+ movq HDR_SIZE_OFFSET(%rax), %rdx
+
+ cmpq $UNAVAILABLE, HDR_AVAIL_OFFSET(%rax) # If the space is
+ # unavailable, i.e.
+ # already in use.. Go
+ je next_location # to the next one.
+
+ cmpq %rdx, %rcx # If the slot is available, check if
+ jle allocate_here # it's big enough.
+
+ next_location:
+ addq $HEADER_SIZE, %rax # Total size of the memory region is the
+ # sum of the size requested (currently
+ # at %rdx), plus the 16 bytes for the
+ addq %rdx, %rax # header.
+
+ jmp alloc_loop_begin # Go look at the next location
+
+ allocate_here: # If we've made it here, that means that
+ # the region header of the region to
+ # allocate is in %rax
+
+ # Mark space as unavailable
+ movq $UNAVAILABLE, HDR_AVAIL_OFFSET(%rax)
+ addq $HEADER_SIZE, %rax # move %rax past the header, so it
+ # points to the usable memory. Such
+ # address is returned to the user.
+
+ # Normal function return
+ movq %rbp, %rsp
+ popq %rbp
+ ret
+
+ move_break: # We have exhausted all addressable memory, we need to
+ # get more from Linux
+ # %rbx holds the current endpoint of the data, and
+ # %rcx its size.
+
+ addq $HEADER_SIZE, %rbx # We need to increase %rbx to where we _want_
+ # memory to end. So we account for the header
+ addq %rcx, %rbx # and the user's requested size
+
+ # Ask Linux for more memory
+
+ # We'll need the values in these registers, so save them...
+ pushq %rax
+ pushq %rbx
+ pushq %rcx
+
+ movq $SYS_BRK, %rax # Set new break by calling brk()
+ movq %rbx, %rdi # x86_64 uses %rdi for first parameter
+ syscall
+
+ # brk() returns 0 for error or the address we asked for (or larger if it
+ # needs to be rounded up). We don't really care here where it ends up
+ # setting the break as long as it isn't 0.
+
+ cmpq $0, %rax # Check for errors
+ je error
+
+ # Restore our registers
+ popq %rcx
+ popq %rbx
+ popq %rax
+
+ # Set this memory as unavailable since we are about to give it away
+ movq $UNAVAILABLE, HDR_AVAIL_OFFSET(%rax)
+ movq %rcx, HDR_SIZE_OFFSET(%rax) # Set memory size
+
+ addq $HEADER_SIZE, %rax # Set %rax to the address being
+ # returned to the user (user
+ # doesn't know anything about the headers
+
+ movq %rbx, current_break # Save the current break. In
+ # reality it may be larger due
+ # rounding, but we don't care
+ # about memory footprint here
+
+ movq %rbp, %rsp
+ popq %rbp
+ ret
+
+ error:
+ movq $0, %rax # Return 0 on error
+ movq %rbp, %rsp
+ popq %rbp
+ ret
+##### END OF allocate ######
+
+## deallocate ##
+#
+# Give back a region of memory to the pool after user is done
+# with it
+#
+# Parameter: Address of the memory to be freed
+# No return value
+#
+# The memory address returned to the user starts at 2 words beyond its header,
+# all we need to do here is mark the memory slot as free (available).
+# By now we don't care about moving the break back.
+.globl deallocate
+.type deallocate, @function
+
+# Stack position of the memory region to be free (function's parameter)
+.equ ST_MEMORY_SEG, 8 # We are not saving %rbp here so we just need to skip ret
+ # address
+
+deallocate:
+ # Function is too simple, no need for fancy function stuff
+
+ # Get the address of the memory to free (Normally this is 16(%rbp), but
+ # since we didn't push %rbp or move %rsp to %rbp, we can just do 8(%rsp)
+ movq ST_MEMORY_SEG(%rsp), %rax
+
+ # Get the pointer to the beginning of the region (i.e. to the header)
+ subq $HEADER_SIZE, %rax
+
+ # Mark it as available
+ movq $AVAILABLE, HDR_AVAIL_OFFSET(%rax)
+ ret
+##### END OF deallocate #####