Char: correct pci_get_device changes
PATCH 1/1 Char: correct pci_get_device changes
correct pci_get_device changes
Commits 881a8c120acf7ec09c90289e2996b7c70f51e996 and efe1ec27837d6639eae82e1f5876910ba6433c3f are totally wrong and "Replace pci_find_device() with more safer pci_get_device()." holds only for a very short time until next iteration of pci_get_device is called (and refcount is decreased again there).
Whenever pci_find_device -> pci_get_device change is performed, also pci_dev_get and pci_dev_put should be in most cases called. If not, it's not easy to find such issues and this is example of such case. Here it was made safer in no way but a moment during initialization, weird.
It affects moxa and rio char drivers. (All this stuff deserves to be converted to pci_probing, though.)
---
commit cf08d779dbf942c5da538e7c286782d05c9da56f
tree 2eaea73812a57c869db26bcc589b4c1b3b2e87f8
parent ed395025a185ff6da5a564a55d320ffd8162304c
author Jiri Slaby <[email protected]> Sun, 15 Oct 2006 01:29:38 +0200 committer Jiri Slaby <[email protected]> Sun, 15 Oct 2006 01:29:38 +0200
drivers/char/moxa.c | 6 ++++++ drivers/char/rio/host.h | 1 + drivers/char/rio/rio_linux.c | 9 +++++++++ 3 files changed, 16 insertions(+), 0 deletions(-)
drivers/char/moxa.c +6 -0
@@ -130,6 +130,7 @@ static moxa_isa_board_conf moxa_isa_boartypedef struct _moxa_pci_devinfo {ushort busNum;ushort devNum; +struct pci_dev *pdev; }moxa_pci_devinfo ; typedef struct _moxa_board_conf {
@@ -324,6 +325,9 @@ static int moxa_get_PCI_conf(struct pci_ board->busType =MOXA_BUS_TYPE_PCI ; board->pciInfo.busNum = p->bus->number; board->pciInfo.devNum = p->devfn >>3 ; + board->pciInfo.pdev = p; +/* don't loss the reference in the next pci_get_device iteration */ +pci_dev_get (p);return (0 ); }
@@ -497,6 +501,8 @@ static void __exit moxa_exit(void)for (i =0 ; i <MAX_BOARDS ; i++) {if (moxaBaseAddr[i])iounmap (moxaBaseAddr[i]); +if (moxa_boards[i].busType ==MOXA_BUS_TYPE_PCI ) +pci_dev_put (moxa_boards[i].pciInfo.pdev); }if (verbose)
drivers/char/rio/host.h +1 -0
@@ -44,6 +44,7 @@ #define MAX_EXTRA_UNITS 64** the host. */ struct Host { +struct pci_dev *pdev;unsigned char Type;/* RIO_EISA, RIO_MCA, ... */ unsigned char Ivec;/* POLLED or ivec number */ unsigned char Mode;/* Control stuff */
drivers/char/rio/rio_linux.c +9 -0
@@ -1017,6 +1017,10 @@ #ifdef CONFIG_PCIrio_dprintk (RIO_DEBUG_PROBE ,"Hmm Tested ok, uniqid = %x.\n" , p->RIOHosts[p->RIONumHosts].UniqueNum);fix_rio_pci (pdev); + + p->RIOHosts[p->RIONumHosts].pdev = pdev; +pci_dev_get (pdev); + p->RIOLastPCISearch =0 ; p->RIONumHosts++; found++;
@@ -1067,6 +1071,9 @@ #ifdef CONFIG_RIO_OLDPCI ((readb (&p->RIOHosts[p->RIONumHosts].Unique[1 ]) &0xFF ) <<8 ) | ((readb (&p->RIOHosts[p->RIONumHosts].Unique[2 ]) &0xFF ) <<16 ) | ((readb (&p->RIOHosts[p->RIONumHosts].Unique[3 ]) &0xFF ) <<24 );rio_dprintk (RIO_DEBUG_PROBE ,"Hmm Tested ok, uniqid = %x.\n" , p->RIOHosts[p->RIONumHosts].UniqueNum); + p->RIOHosts[p->RIONumHosts].pdev = pdev; +pci_dev_get (pdev); + p->RIOLastPCISearch =0 ; p->RIONumHosts++; found++;
@@ -1188,6 +1195,8 @@ static void __exit rio_exit(void)if (hp->Caddr)iounmap (hp->Caddr); +if (hp->Type ==RIO_PCI ) +pci_dev_put (hp->pdev); }if (misc_deregister (&rio_fw_device) <0 ) {
Ar Sul, 2006-10-15 am 01:36 +0200, ysgrifennodd Jiri Slaby:
Quoted 4 lines
correct
pci_get_devicechangesCommits
881a8c120acf7ec09c90289e2996b7c70f51e996andefe1ec27837d6639eae82e1f5876910ba6433c3fare totally wrong and
No I disagree, they are totally right. They stop crashes walking the list during a hotplug of another card. If you hotplug the moxa or rio card you are screwed. Your patches solve a different problem to the ones the pci_dev_get changes solve.
Your changes also minimally handle the hotplug of the moxa or rio card which I guess could theoretically happen in some weird system and are a definite improvement.
Quoted 2 lines
It affects moxa and rio char drivers. (All this stuff deserves to be converted to
pci_probing, though.)
Agreed, or dropped
Quoted 13 lines · Jiri Slaby
correct
pci_get_devicechangesCommits
881a8c120acf7ec09c90289e2996b7c70f51e996andefe1ec27837d6639eae82e1f5876910ba6433c3fare totally wrong and "Replacepci_find_device()with more saferpci_get_device()." holds only for a very short time until next iteration ofpci_get_deviceis called (and refcount is decreased again there).Whenever
pci_find_device->pci_get_devicechange is performed, alsopci_dev_getandpci_dev_putshould be in most cases called. If not, it's not easy to find such issues and this is example of such case. Here it was made safer in no way but a moment during initialization, weird.
Quoted 7 lines
It affects moxa and rio char drivers. (All this stuff deserves to be converted to
pci_probing, though.)Cc <[email protected]>Cc Amit Gud <[email protected]>Cc Greg Kroah-Hartman <[email protected]>Signed-off-by Jiri Slaby <[email protected]>
Quoted 1 line · On Sun, 15 Oct 2006 01:36:45 +0200 (CEST) Jiri Slaby <[email protected]>…
correct
pci_get_devicechanges
I suspect this patch was against -mm, but fixes problems in mainline, to which it doesn't apply.
Could I have a new one against mainline please? With a changelog updated to reflect Alan's comments?
Thanks.
Quoted 10 lines · Andrew Morton
Quoted 1 line · On Sun, 15 Oct 2006 01:36:45 +0200 (CEST) Jiri Slaby <[email protected]>…
correct
pci_get_devicechangesI suspect this patch was against -mm, but fixes problems in mainline, to which it doesn't apply.
Could I have a new one against mainline please? With a changelog updated to reflect Alan's comments?
Sure, thanks,
Signature · 4 lines
http://www.fi.muni.cz/~xslaby/ Jiri Slaby faculty of informatics, masaryk university, brno, cz e-mail: jirislaby gmail com, gpg pubkey fingerprint: B674 9967 0407 CE62 ACC8 22A0 32CC 55C3 39D4 7A7E