perf: dont use recursion for build_prp_list + remove redundant calls

drivers/nvme/nvme.asm

@@ -452,7 +452,7 @@ proc get_log_page stdcall, pci:dword, dptr:dword, lid:byte
 
 endp
 
-proc free_prp_list stdcall, prp_list_ptr:dword, nprps:dword, recursive:byte
+proc free_prp_list stdcall, prp_list_ptr:dword, nprps:dword
 
 	push 	edi
 	mov 	edi, [prp_list_ptr]
@@ -482,69 +482,96 @@ endp
 proc build_prp_list stdcall, nprps:dword
 
 	push 	esi ebx edi
+	sub 	esp, 16
+
+	; here, we store the pointer to the very first
+	; PRP list so that free_prp_list can free the
+	; entire PRP list if something goes wrong, it
+	; also serves as our return value placeholder
+	mov 	dword [esp], 0
+
+	; store the number of PRPs here, we need to 
+	; cache the result here since we'll subtract
+	; nprps regularly. If we don't we will lose
+	; the value.
+	mov 	eax, [nprps]
+	mov 	dword [esp + 4], eax
+
+	; store consecutive PRP list buffer here, since
+	; we'll increment the PRP list pointer continuously
+	; its important to cache the value
+	mov 	dword [esp + 12], 0
+
 	xor 	edi, edi
 	xor 	esi, esi
 	mov 	ecx, [nprps]
 	mov 	ebx, ecx
 	shl 	ecx, 3
+	; we'll store consecutive PRP list buffers here, for example
+	; given 2 PRP lists, we allocate 2 continuous pages
 	invoke  KernelAlloc, ecx ; store pointers to the PRP entries here
 	test 	eax, eax
 	jz 	.err
+	mov 	dword [esp + 12], eax
 	mov 	edi, eax
+
+.build_prp_list:
 	mov 	ecx, ebx
+	cmp 	ecx, PAGE_SIZE / 8
+	jne 	@f
+	mov 	ecx, PAGE_SIZE / 8
+
+@@:
 	imul 	ecx, PAGE_SIZE
+
+	; allocate a massive buffer for the PRPs themselves
 	invoke  KernelAlloc, ecx
 	test 	eax, eax
 	jz 	.err
+	invoke  GetPhysAddr
 	mov 	esi, eax
 	xor 	ecx, ecx
-	push 	esi
+	cmp 	dword [esp], 0
+	jz 	@f
 
-.loop:
+	; we need to store the pointer of the newly allocated 
-	; check if we need to build another PRP list
+	; PRP list to the previous PRP list last slot
-	mov 	eax, ebx
+	mov 	eax, dword [esp + 8]
-	sub 	eax, PAGE_SIZE
+	mov 	dword [eax + PAGE_SIZE - 8], esi
-	; is this the last entry in the list? If so, build yet another PRP list here
+	mov 	dword [eax + PAGE_SIZE - 4], 0
-	cmp 	ecx, PAGE_SIZE / 8 - 1
-	jne 	@f
-	push 	ecx
-	stdcall build_prp_list, eax, TRUE
-	pop 	ecx
-	test 	eax, eax
-	jz 	.cleanup_prp_list
-	invoke  GetPhysAddr
-	mov 	dword [edi + ecx * 4], eax
-	mov 	dword [edi + ecx * 4 + 4], 0
-	jmp 	.success
 
 @@:
-	mov 	eax, esi
+	mov 	dword [esp], esi
-	push 	ecx
+
-	invoke  GetPhysAddr
+.loop:
-	pop 	ecx
+	mov 	dword [edi + ecx * 4], esi
-	mov 	dword [edi + ecx * 4], eax
 	mov 	dword [edi + ecx * 4 + 4], 0
 	add 	esi, PAGE_SIZE
 	inc 	ecx
 	cmp 	ecx, ebx
-	jb 	.loop
+	jne 	.loop
-
+	; check if we we need to build another PRP list
-.success:
+	mov 	dword [esp + 8], edi
-	mov 	eax, edi
+	mov 	ebx, [nprps]
-	pop 	esi
+	cmp 	ebx, PAGE_SIZE / 8
-	pop 	edi ebx esi
+	je 	.build_prp_list
-	ret
-
-.cleanup_prp_list:
-	pop 	esi
-	invoke  KernelFree, esi
 
 .err:
-	test 	edi, edi
+	mov 	esi, dword [esp]
+	test 	esi, esi
 	jz 	@f
-	invoke  KernelFree, edi
+	sub 	[nprps], PAGE_SIZE / 8
+	add 	edi, PAGE_SIZE
+	stdcall free_prp_list, esi, [esp + 4]
 
 @@:
+	mov 	eax, dword [esp + 12]
+	test 	eax, eax
+	jz 	@f
+	invoke  KernelFree, eax
+
+@@:
+	add 	esp, 16
 	pop 	edi ebx esi
 	xor 	eax, eax
 	ret
@@ -575,7 +602,6 @@ proc alloc_dptr stdcall, ns:dword, prps_ptr:dword, numsectors:dword
 	cmp 	edx, eax
 	jae 	@f
 	
-	DEBUGF  DBG_INFO, "Allocating single entry for PRP2\n"
 	; allocate a single PRP entry for PRP2
 	invoke  AllocPage
 	test 	eax, eax