some refactoring and fixes

drivers/nvme/nvme.asm

@@ -450,120 +450,122 @@ proc get_log_page stdcall, pci:dword, dptr:dword, lid:byte
 
 endp
 
+proc free_prp_list stdcall, prp_list_ptr:dword, nprps:dword
+
+		push 	ebx edi
+		mov 	edi, [prp_list_ptr]
+		xor 	ebx, ebx
+
+@@:
+		invoke  KernelFree, dword [edi + ebx * 4]
+		inc 	ebx
+		cmp 	ebx, [nprps]
+		jb 		@b
+		pop 	edi ebx
+		xor 	eax, eax
+		ret
+
+endp
+
 proc build_prp_list stdcall, nprps:dword
 
 		push 	ebx edi
-		mov 	ebx, [nprps]
+		xor 	edi, edi
-		push 	ebx
+		mov 	ecx, [nprps]
-		shl 	ebx, 2
+		mov 	ebx, ecx
-		invoke  KernelAlloc, ebx
+		shl 	ecx, 3
-		pop 	ebx
+		invoke  KernelAlloc, ecx ; store pointers to the PRP entries here
 		test 	eax, eax
 		jz 		.err
 		mov 	edi, eax
-		xor 	ecx, ecx
+		mov 	ecx, [nprps]
-		
+		imul 	ecx, PAGE_SIZE
-@@:
+		invoke  KernelAlloc, ecx
-		invoke  KernelAlloc, 0x1000
 		test 	eax, eax
-		jz 		.cleanup
+		jz 		.err
-		mov 	dword [edi + ecx], eax
+		mov 	edx, eax
+		xor 	ecx, ecx
+
+@@:
+		push 	edx ecx
+		invoke  GetPhysAddr, edx
+		pop 	ecx edx
+		mov 	dword [edi + ecx * 4], eax
+		mov 	dword [edi + ecx * 4 + 4], 0
+		add 	edx, PAGE_SIZE
 		inc 	ecx
 		cmp 	ecx, ebx
 		jb 		@b
-		mov 	dword [esi + ecx], edi
-		mov 	eax, edi
 		pop 	edi ebx
 		ret
-
+		
-.cleanup:
+.err:
-		xor 	ecx, ecx
+		test 	edi, edi
+		jnz 	@f
+		invoke  KernelFree, edi
 
 @@:
-		invoke  KernelFree, dword [edi + ecx]
-		inc 	ecx
-		cmp 	ecx, edx
-		jb 		@b
-
-.err:
-		xor 	eax, eax
 		pop 	edi ebx
+		xor 	eax, eax
 		ret
 		
 endp
 
-; Returns a PRP/PRP List
+proc alloc_dptr stdcall, ns:dword, prps_ptr:dword, start_sector:qword, numsectors:dword
-; ----------------------------------------------------------------------------------------------------
-; flags (used for identifying how to free the memory used by the PRPs after the command is completed):
-; BIT 1 SET: allocated PRP1 entry
-; BIT 2 SET: allocated PRP1 list
-; BIT 3 SET: allocated PRP2 entry
-; BIT 4 SET: allocated PRP2 list
-proc alloc_dptr stdcall, ns:dword, prps_ptr:dword, start_sector:qword, numsectors:dword, flags_ptr:dword
 
-		push 	esi edi ebx
+		push 	esi edi
 		mov 	esi, [ns]
-		mov 	edi, [flags_ptr]
+		mov 	edi, [prps_ptr]
-		mov 	ebx, dword [esi + NSINFO.pg_sectors]
+		mov 	qword [esi], 0
-		mov 	esi, [prps_ptr]
+		invoke  KernelAlloc, PAGE_SIZE
-		mov 	dword [esi + 4], 0 ; set PRP #2 to reserved by default
-		cmp 	ebx, dword [edi]
-		ja 		.numsectors_over_1pg
-
-.alloc_prp1_entry:
-		invoke 	KernelAlloc, 0x1000
 		test 	eax, eax
 		jz 		.err
-		or 		byte [edi], PRP1_ENTRY_ALLOCATED
+		mov 	dword [edi], eax
-		mov 	dword [esi], eax
-		pop 	ebx edi esi
-		ret
 
-.numsectors_over_1pg:
+		; is the number of sectors equal to two memory pages in size?
-		; check if offset portion of PRP1 is 0
+		mov 	eax, dword [esi + NSINFO.pg_sectors]
-		mov 	eax, dword [esi]
-		and 	eax, (PAGE_SIZE shl NVM_MPS) - 1
-		test 	eax, eax
-		jz 		.check_prp2_cond_bii
-
-		; check PRP2 condition b) i.
-		cmp 	ebx, [numsectors]
-		jne 	.check_prp2_cond_cici
-		invoke  KernelAlloc, 0x1000
-		test 	eax, eax
-		jz 		.err
-		mov 	dword [esi + 4], eax
-		or 		byte [edi], PRP2_ENTRY_ALLOCATED
-		jmp 	.alloc_prp1_entry
-		
-		
-.check_prp2_cond_bii:
-		mov 	eax, ebx
 		shl 	eax, 1
 		cmp 	eax, [numsectors]
-		ja 		.check_prp2_cond_c
+		jne 	@f
-		invoke  KernelAlloc, 0x1000
-		test 	eax, eax
-		jz 		.err
-		mov 	dword [esi + 4], eax
-		or 		byte [edi], PRP2_ENTRY_ALLOCATED
-		jmp 	.alloc_prp1_entry
-
-.check_prp2_cond_c:
-		; is the command data transfer length greater or equal to two memory pages?
-		mov 	eax, ebx
-		shl 	eax, 1
-		cmp 	eax, [numsectors_ptr]
-		jb  	.alloc_prp1_entry
-		stdcall build_prp_list, 4 ; TODO: replace 4 with proper number of PRPs later
-		test 	eax, eax
-		jz 		.err
-		mov 	dword [esi + 4], eax
-		jmp 	.alloc_prp1_entry
 		
+		; allocate a single PRP entry for PRP2
+		invoke  KernelAlloc, PAGE_SIZE
+		test 	eax, eax
+		jz 		.err
+		mov 	dword [edi + 4], eax
+		jmp 	.success
+
+@@:
+		; allocate PRP list for PRP2
+		shr 	eax, 1
+		mov 	ecx, [numsectors]
+		xchg 	eax, ecx
+		xor 	edx, edx
+		div 	ecx
+		stdcall build_prp_list, eax
+		test 	eax, eax
+		jz 		.free_prp_list
+		mov 	dword [edi + 4], eax
+		jmp 	.success
+		
+.free_prp_list:
+		invoke  KernelFree, eax
+
 .err:
+		mov 	eax, dword [edi]
+		test 	eax, eax
+		jz 		@f
+		invoke  KernelFree, eax
+
+@@:
 		xor 	eax, eax
-		pop 	ebx edi esi
+		pop 	edi esi
+		ret
+
+.success:
+		xor 	eax, eax
+		inc 	eax
+		pop 	edi esi
 		ret
 		
 endp
@@ -830,6 +832,7 @@ proc nvme_init stdcall, pci:dword
 		stdcall memset, eax, 0, sizeof.NVM_QUEUE_ENTRY * (LAST_QUEUE_ID + 1)
 
 		; Allocate submission/completion queue pointers
+		; TODO: Make these queues physically contiguous
 		xor 	ecx, ecx
 @@:
 		push 	ecx
@@ -907,6 +910,7 @@ proc nvme_init stdcall, pci:dword
 		DEBUGF  DBG_INFO, "nvme%u: I/O controller detected...\n", [esi + pcidev.num]
 
 @@:
+		; TODO: check IDENTC.AVSCC
 		mov 	al, byte [edi + IDENTC.sqes]
 		and 	al, 11110000b
 		cmp 	al, 0x60 ; maximum submission queue size should at least be 64 bytes

drivers/nvme/nvme.inc

@@ -16,7 +16,6 @@ VS121 = 0x00010201 ; (v1.2.1)
 VS130 = 0x00010300 ; (v1.3.0)
 VS140 = 0x00010400 ; (v1.4.0)
 
-PAGE_SIZE = 4096
 NVM_CMDS = 64       ; Number of Commands
 NVM_MPS = 0         ; Memory Page Size (2 ^ (12 + MPS))
 NVM_ASQS = 63        ; Admin Submission Queue Size
@@ -24,6 +23,7 @@ NVM_ACQS = NVM_ASQS ; Admin Completion Queue Size
 LAST_QUEUE_ID = 1 ; Index of the last queue
 SQ_ENTRIES = NVM_ASQS ; I/O and Admin Submission Queue Size
 CQ_ENTRIES = NVM_ACQS ; I/O and Admin Completion Queue Size
+PAGE_SIZE = 4096 shl NVM_MPS
 
 ADMIN_QUEUE = 0 ; Admin Queue ID