On Wed, Mar 01, 2017 at 12:17:02AM +0900, YASUOKA Masahiko wrote: > Hi, > > On Tue, 28 Feb 2017 15:16:58 +0100 > Patrick Wildt <[email protected]> wrote: > > On Tue, Feb 28, 2017 at 10:41:21PM +0900, YASUOKA Masahiko wrote: > >> On Mon, 27 Feb 2017 11:10:31 +0100 > >> Patrick Wildt <[email protected]> wrote: > >> > I'm surprised this didn't come up earlier, but I think the for-loop is a > >> > bit wrong. What the code is supposed to be doing is going over each > >> > device path node for the loaded image, which is supposed to be the path > >> > to the device that efiboot was loaded from, and check for a node that > >> > is a medium (as in hard drive, cdrom) and look for a hard drive. > >> > >> Yes, you are right it's wrong. > >> > >> > As it turns out, the code would actually skip all device paths that > >> > are mediums and instead match on all non-media types. It's surprising > >> > to me that this seems to work for people. > >> > > >> > On Qemu and VMware the first node in the device path is ACPI (2) with > >> > subtype (1). Interestingly, the hard drive subtype in the media type > >> > has the subtype (1) as well. So that might be a reason it did work. > >> > >> As far as my test on QEMU and VAIO, the first node is ACPI and subtype > >> is 1. > >> > >> But we had another bug, > >> > >> 87 if (status == EFI_SUCCESS) { > >> 88 for (dp = dp0; !IsDevicePathEnd(dp); > >> 89 dp = NextDevicePathNode(dp)) { > >> 90 if (DevicePathType(dp) == > >> MEDIA_DEVICE_PATH) > >> 91 continue; > >> 92 if (DevicePathSubType(dp) == > >> MEDIA_HARDDRIVE_DP) { > >> 93 bios_bootdev = 0x80; > >> 94 efi_bootdp = dp; > >> 95 break; > >> 96 } > >> 97 break; > >> 98 } > >> 99 } > >> > >> if we fix the bug at #90, at #94 efi_bootdp will not become the first > >> dp node(dp0). Then efi_diskprobe() will fail to reorder hd? properly > >> since the function assumes efi_bootdp is the first dp node. > >> > >> So I'd like to commit the diff below. > > > > First of all, I've been having a look at supporting CDROMs in efiboot > > and saw that FreeBSD has a nice "devpath.c" file which implements a few > > good helpers. I've been using efi_devpath_match(), efi_devpath_trim() > > and efi_devpath_last_node(). Those have been particularly helpful and > > simplify some code. I wonder if it would make sense to pull that one > > in instead of writing another completely new implementation of tha same > > thing. > > > >> diff --git a/sys/arch/amd64/stand/efiboot/efiboot.c > >> b/sys/arch/amd64/stand/efiboot/efiboot.c > >> index 1903862d271..f82ea34f66e 100644 > >> --- a/sys/arch/amd64/stand/efiboot/efiboot.c > >> +++ b/sys/arch/amd64/stand/efiboot/efiboot.c > >> @@ -41,7 +41,8 @@ > >> EFI_SYSTEM_TABLE *ST; > >> EFI_BOOT_SERVICES *BS; > >> EFI_RUNTIME_SERVICES *RS; > >> -EFI_HANDLE IH, efi_bootdp = NULL; > >> +EFI_HANDLE IH; > >> +EFI_DEVICE_PATH *efi_bootdp = NULL; > >> EFI_PHYSICAL_ADDRESS heap; > >> EFI_LOADED_IMAGE *loadedImage; > >> UINTN heapsiz = 1 * 1024 * 1024; > >> @@ -51,6 +52,7 @@ static EFI_GUID blkio_guid = BLOCK_IO_PROTOCOL; > >> static EFI_GUID devp_guid = DEVICE_PATH_PROTOCOL; > >> u_long efi_loadaddr; > >> > >> +static int efi_device_path_cmp(EFI_DEVICE_PATH *, EFI_DEVICE_PATH > >> *, int); > >> static void efi_heap_init(void); > >> static void efi_memprobe_internal(void); > >> static void efi_video_init(void); > >> @@ -87,14 +89,12 @@ efi_main(EFI_HANDLE image, EFI_SYSTEM_TABLE *systab) > >> if (status == EFI_SUCCESS) { > >> for (dp = dp0; !IsDevicePathEnd(dp); > >> dp = NextDevicePathNode(dp)) { > >> - if (DevicePathType(dp) == MEDIA_DEVICE_PATH) > >> - continue; > >> - if (DevicePathSubType(dp) == MEDIA_HARDDRIVE_DP) { > >> + if (DevicePathType(dp) == MEDIA_DEVICE_PATH && > >> + DevicePathSubType(dp) == MEDIA_HARDDRIVE_DP) { > >> bios_bootdev = 0x80; > >> - efi_bootdp = dp; > >> + efi_bootdp = dp0; > > > > This part makes sense to me. > > > >> break; > >> } > >> - break; > >> } > >> } > >> > >> @@ -166,7 +166,7 @@ efi_diskprobe(void) > >> EFI_BLOCK_IO *blkio; > >> EFI_BLOCK_IO_MEDIA *media; > >> struct diskinfo *di; > >> - EFI_DEVICE_PATH *dp, *bp; > >> + EFI_DEVICE_PATH *dp; > >> > >> TAILQ_INIT(&efi_disklist); > >> > >> @@ -193,23 +193,14 @@ efi_diskprobe(void) > >> di = alloc(sizeof(struct diskinfo)); > >> efid_init(di, blkio); > >> > >> - if (efi_bootdp == NULL) > >> - goto next; > >> - status = EFI_CALL(BS->HandleProtocol, handles[i], &devp_guid, > >> - (void **)&dp); > >> - if (EFI_ERROR(status)) > >> - goto next; > >> - bp = efi_bootdp; > >> - while (1) { > >> - if (IsDevicePathEnd(dp)) { > >> + if (efi_bootdp != NULL) { > >> + status = EFI_CALL(BS->HandleProtocol, handles[i], > >> &devp_guid, > >> + (void **)&dp); > >> + if (EFI_ERROR(status)) > >> + goto next; > >> + if (efi_device_path_cmp(efi_bootdp, dp, > >> MESSAGING_DEVICE_PATH) > >> + == 0) > > > > This part makes sense to me as well. > > > >> bootdev = 1; > >> - break; > >> - } > >> - if (memcmp(dp, bp, sizeof(EFI_DEVICE_PATH)) != 0 || > >> - memcmp(dp, bp, DevicePathNodeLength(dp)) != 0) > >> - break; > >> - dp = NextDevicePathNode(dp); > >> - bp = NextDevicePathNode(bp); > >> } > >> next: > >> if (bootdev) > >> @@ -221,6 +212,35 @@ next: > >> free(handles, sz); > >> } > >> > >> +static int > >> +efi_device_path_cmp(EFI_DEVICE_PATH *dpa, EFI_DEVICE_PATH *dpb, int > >> dptype) > >> +{ > >> + int cmp; > >> + EFI_DEVICE_PATH *dp, *dpt_a = NULL, *dpt_b = NULL; > >> + > >> + for (dp = dpa; !IsDevicePathEnd(dp); dp = NextDevicePathNode(dp)) { > >> + if (DevicePathType(dp) == dptype) { > >> + dpt_a = dp; > >> + break; > >> + } > >> + } > >> + for (dp = dpb; !IsDevicePathEnd(dp); dp = NextDevicePathNode(dp)) { > >> + if (DevicePathType(dp) == dptype) { > >> + dpt_b = dp; > >> + break; > >> + } > >> + } > >> + > >> + if (dpt_a && dpt_b) { > >> + cmp = DevicePathNodeLength(dpt_a) - DevicePathNodeLength(dpt_b); > >> + if (cmp) > >> + return (cmp); > >> + return (memcmp(dpt_a, dpt_b, DevicePathNodeLength(dpt_a))); > >> + } > >> + > >> + return ((uintptr_t)dpt_a - (uintptr_t)dpt_b); > >> +} > > > > I don't like this. Have a look at the FreeBSD implementation, it's > > nicer sind you don't have to pass some device path type but instead > > it does a full path comparison. > > How will you fix by efi_devpath_match()? > > AFAIK, we need to find the device path of the booted disk. (not the > booted parition) So I belive "full path comparison" is not usable. > > --yasuoka >
I see the problem. The way people do it with hard drives and cdroms is to trim/cut away the last node of the device path, because then you get the device path to the entire device instead of just a subnode. Then you can compare the pathes.
