Skip to contentSkip to conversation

ide-cd: prevent null pointer deref via cdrom_newpc_intr

PATCH In review Rainer Weikusat 3 participants 11 messages Jun 18, 2009 – Jun 20, 2009 libata

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ide-cd: prevent null pointer deref via cdrom_newpc_intr

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. This doesn't work for requests which don't have bios associated with them ('GPCMD_READ_DISC_INFO'), because the first call to ide_end_rq, done via ide_complete_rq in order to do the partial completion part, returns with a code of zero for all non-bio requests, causing the drive->hwif->rq pointer to be set to NULL. Upon calling ide_complete_rq a second time, it is attempted to de-reference this null pointer, resulting in a kernel crash.

---

This is fixed in the linux-ide tree since at about 2009/06/10 [Bug 13399, also happens w/ TSSTcorpDVD-ROM SH-D162C], but a patch against 2.6.30 AFAIK doesn't exist (and I didn't find the corresponding thread before digging through all of this ...).

ide/ide-cd.c renamed from ide/ide-cd.c.orig +1 -1
@@ -758,7 +758,7 @@ out_end:
 				rq->errors = -EIO;
 		}
 
-		if (uptodate == 0)
+		if (uptodate == 0 && rq->bio)
 			ide_cd_error_cmd(drive, cmd);
 
 		/* make sure it's fully ended */
#1 lore raw

Hi,

we we're just debugging this and I was about to send the patches tomorrow but you beat me to it :).

Quoted 11 lines · Rainer Weikusat

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. This doesn't work for requests which don't have bios associated with them ('GPCMD_READ_DISC_INFO'), because the first call to ide_end_rq, done via ide_complete_rq in order to do the partial completion part, returns with a code of zero for all non-bio requests, causing the drive->hwif->rq pointer to be set to NULL.

This is a bit misleading, it should be more like: "ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq but the rq is without a bio and the block layer does partial completion only for requests with bio's so this request is completed as a whole and the rq freed."

please fix.

Quoted 10 lines

Upon
calling ide_complete_rq a second time, it is attempted to de-reference this null pointer, resulting in a kernel crash.

---

This is fixed in the linux-ide tree since at about 2009/06/10 [Bug 13399, also happens w/ TSSTcorpDVD-ROM SH-D162C],

really, because I can't find it in Bart's trees. Do you have a commit id?

Quoted 15 lines

but a patch
against 2.6.30 AFAIK doesn't exist (and I didn't find the corresponding thread before digging through all of this ...).

ide/ide-cd.c        2009-06-18 14:10:16.000000000 +0200 renamed from ide/ide-cd.c.orig   2009-06-18 15:10:24.000000000 +0200 +1 -1
@@ -758,7 +758,7 @@ out_end:
                                rq->errors = -EIO;
                }
 
-               if (uptodate == 0)
+               if (uptodate == 0 && rq->bio)
                        ide_cd_error_cmd(drive, cmd);
 
                /* make sure it's fully ended */
Signature · 2 lines
Regards/Gruss,
Boris
#2 lore raw
Quoted 13 lines · Borislav Petkov
Quoted 11 lines · Rainer Weikusat

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. This doesn't work for requests which don't have bios associated with them ('GPCMD_READ_DISC_INFO'), because the first call to ide_end_rq, done via ide_complete_rq in order to do the partial completion part, returns with a code of zero for all non-bio requests, causing the drive->hwif->rq pointer to be set to NULL.

Quoted 5 lines

This is a bit misleading, it should be more like: "ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq but the rq is without a bio and the block layer does partial completion only for requests with bio's so this request is completed as a whole and the rq freed."

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

Quoted 1 line

please fix.

I will send a modified 'patch e-mail' soon.

Something I would like to add: The DVD-ROM mentioned below has exactly the same 32/30 issue w/ READ DISC INFO. This used to just be an unnoticed failure in older kernels.

[...]

Quoted 5 lines
Quoted 2 lines

This is fixed in the linux-ide tree since at about 2009/06/10 [Bug 13399, also happens w/ TSSTcorpDVD-ROM SH-D162C],

really, because I can't find it in Bart's trees. Do you have a commit id?

No, I just assumed that, since I found the bio-check among beginnings of code intended to deal with the 32/30 issue.

#3 lore raw

Hi,

Quoted 14 lines · Rainer Weikusat
Quoted 13 lines · Borislav Petkov
Quoted 11 lines · Rainer Weikusat

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. This doesn't work for requests which don't have bios associated with them ('GPCMD_READ_DISC_INFO'), because the first call to ide_end_rq, done via ide_complete_rq in order to do the partial completion part, returns with a code of zero for all non-bio requests, causing the drive->hwif->rq pointer to be set to NULL.

Quoted 8 lines
Quoted 5 lines

This is a bit misleading, it should be more like: "ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq but the rq is without a bio and the block layer does partial completion only for requests with bio's so this request is completed as a whole and the rq freed."

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

That doesn't matter because the OOPS happens after the command has been issued and _before_ ide_cd_queue_pc() gets to access the rq ptr again.

ide_cd_queue_pc() does blk_execute_rq() which does wait for completion of the request so it doesn't access the rq pointer before the IRQ handler has completed. Even if it would reach the blk_put_request part, the OOPS would have already happened.

ide_complete_rq simply does

blk_end_request
|->blk_end_bidi_request
|->blk_finish_request

after checking the rq->bio pointer through blk_update_bidi_request() and if the prior is NULL it does __blk_put_request in blk_finish_request and this is where the thing vanishes.

The second time ide_complete_rq() comes to run at the end of the IRQ handler the rq is already 0xdeadbeef :).

(this is all latest git but the old code (without the bidi stuff) did the same thing wrt to rq->bio).

Signature · 2 lines
Regards/Gruss,
Boris
#4 lore raw
Quoted 2 lines · Borislav Petkov
Quoted 1 line · Rainer Weikusat

Borislav Petkov <[email protected]> writes:

[...]

Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

so this request is completed as a whole and the rq freed."

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

That doesn't matter because the OOPS happens after the command has been issued and _before_ ide_cd_queue_pc() gets to access the rq ptr again.

Yes. Because the pointer I already mentioned has been reset. But the request itself is still alive.

[...]

Quoted 6 lines

ide_complete_rq simply does

blk_end_request
|->blk_end_bidi_request
|->blk_finish_request

Quoted 6 lines

after checking the rq->bio pointer through blk_update_bidi_request() and if the prior is NULL it does __blk_put_request in blk_finish_request and this is where the thing vanishes.

The second time ide_complete_rq() comes to run at the end of the IRQ handler the rq is already 0xdeadbeef :).

Not quite. Below is the blk_execute_rq-function:

int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk,
struct request *rq, int at_head)

{
        DECLARE_COMPLETION_ONSTACK(wait);
        char sense[SCSI_SENSE_BUFFERSIZE];
        int err = 0;

        /*
         * we need an extra reference to the request, so we can look at
         * it after io completion
         */
        rq->ref_count++;

        if (!rq->sense) {
                memset(sense, 0, sizeof(sense));
                rq->sense = sense;
                rq->sense_len = 0;
        }

        rq->end_io_data = &wait;
        blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq);
        wait_for_completion(&wait);

        if (rq->errors)
                err = -EIO;

        return err;
}

and the refcount is incremented at the start of that 'so we can look at it after io-completion', which means 'after the code below has executed':

static void blk_end_sync_rq(struct request *rq, int error)

{
        struct completion *waiting = rq->end_io_data;

        rq->end_io_data = NULL;
        __blk_put_request(rq->q, rq);

        /*
         * complete last, if this is a stack request the process (and thus
         * the rq pointer) could be invalid right after this complete()
         */
        complete(waiting);
}

which puts the rq once, decrementing the refcount by
one. Both blk_execute_rq and ide_cd_queue_pc inspect the rq after it has been completed and the latter puts it again. While inspecting a 'freed' data structure would probably be harmless, freeing it twice would be quite of a problem :-). I have spent something like 18 hours of my life to determine what was going on here, starting from 'I have never seen any of this again' and came across a bit of it in the process.

But I am actually busy doing 'something completely different' with 2.4 for my job ...

#5 lore raw
Quoted 83 lines · Rainer Weikusat
Quoted 2 lines · Borislav Petkov
Quoted 1 line · Rainer Weikusat

Borislav Petkov <[email protected]> writes:

[...]

Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

so this request is completed as a whole and the rq freed."

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

That doesn't matter because the OOPS happens after the command has been issued and _before_ ide_cd_queue_pc() gets to access the rq ptr again.

Yes. Because the pointer I already mentioned has been reset. But the request itself is still alive.

[...]

Quoted 6 lines

ide_complete_rq simply does

blk_end_request
|->blk_end_bidi_request
   |->blk_finish_request

Quoted 6 lines

after checking the rq->bio pointer through blk_update_bidi_request() and if the prior is NULL it does __blk_put_request in blk_finish_request and this is where the thing vanishes.

The second time ide_complete_rq() comes to run at the end of the IRQ handler the rq is already 0xdeadbeef :).

Not quite. Below is the blk_execute_rq-function:

int blk_execute_rq(struct request_queue *q, struct gendisk *bd_disk, struct request *rq, int at_head)
{
       DECLARE_COMPLETION_ONSTACK(wait);
       char sense[SCSI_SENSE_BUFFERSIZE];
       int err = 0;

       /*
        * we need an extra reference to the request, so we can look at
        * it after io completion
        */
       rq->ref_count++;

       if (!rq->sense) {
               memset(sense, 0, sizeof(sense));
               rq->sense = sense;
               rq->sense_len = 0;
       }

       rq->end_io_data = &wait;
       blk_execute_rq_nowait(q, bd_disk, rq, at_head, blk_end_sync_rq); wait_for_completion(&wait);

       if (rq->errors)
               err = -EIO;

       return err;
}

and the refcount is incremented at the start of that 'so we can look at it after io-completion', which means 'after the code below has executed':

static void blk_end_sync_rq(struct request *rq, int error) {
       struct completion *waiting = rq->end_io_data;

       rq->end_io_data = NULL;
       __blk_put_request(rq->q, rq);

       /*
        * complete last, if this is a stack request the process (and thus
        * the rq pointer) could be invalid right after this complete() */
       complete(waiting);
}

Quoted 2 lines

which puts the rq once, decrementing the refcount by one.

Please, look at the following trace:

ide-cd: ide_cd_queue_pc: cmd[0]: 0x51, write: 0x0, timeout: 1750,
cmd_flags: 0x8000
ide-cd: ide_cd_do_request: cmd: 0x51, block: 18446744073709551615
ide_cd_do_request: dev hda: type=a, flags=108a640
sector 18446744073709551615, nr/cnr 0/0
bio (null), biotail (null), buffer (null), data ffff88011b849ba0, len 32
ide-cd: cdrom_do_block_pc: rq->cmd[0]: 0x51, rq->cmd_type: 0xa
ide-cd: cdrom_newpc_intr: cmd: 0x51, write: 0x0
ide-cd: cdrom_newpc_intr: DRQ: stat: 0x58, thislen: 30
ide-cd: ide_cd_check_ireason: ireason: 0x2, rw: 0x0
ide-cd: cdrom_newpc_intr: data transfer, rq->cmd_type: 0xa, ireason: 0x2
ide-cd: cdrom_newpc_intr: cmd: 0x51, write: 0x0
ide-cd: cdrom_newpc_intr: DRQ: stat: 0x50, thislen: 2
ide-cd: ide_cd_request_sense_fixup: rq->cmd[0]: 0x51
#3

[ this is a printk I added to show from where we hit the out_end label. There we call the first ide_complete_rq() over ide_cd_error_cmd() where we put the request. __blk_put_request returns without freeing the rq if the refcount is > 1, which, in this case, is.]

and now here we do the second direct ide_complete_rq and here the rq is freed:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000048

However, despite the refcounting, the rq->bio check kills the rq - otherwise the block layer would've waited, see the beginning of blk_update_request():

	if (!req->bio)
		return false;

but let me do more tracing to see exactly what happens and when it happens, now that we're waist-deep into the block layer :).

Signature · 2 lines
Regards/Gruss,
Boris
#6 lore raw
Quoted 10 lines · Borislav Petkov
Quoted 9 lines · Rainer Weikusat
Quoted 2 lines · Borislav Petkov
Quoted 1 line · Rainer Weikusat

Borislav Petkov <[email protected]> writes:

[...]

Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

so this request is completed as a whole and the rq freed."

Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

That doesn't matter because the OOPS happens after the command has been issued and _before_ ide_cd_queue_pc() gets to access the rq ptr again.

Yes. Because the pointer I already mentioned has been reset.

And this happens here (!!!, code from 2.6.30 vanilla):

int ide_complete_rq(ide_drive_t *drive, int error, unsigned int nr_bytes)

{
	ide_hwif_t *hwif = drive->hwif;
	struct request *rq = hwif->rq;
	int rc;

	/*
	 * if failfast is set on a request, override number of sectors
	 * and complete the whole request right now
	 */
	if (blk_noretry_request(rq) && error <= 0)
		nr_bytes = rq->hard_nr_sectors << 9;

	rc = ide_end_rq(drive, rq, error, nr_bytes); 
	if (rc == 0)
		hwif->rq = NULL; /* !!! */

	return rc;
}

[explanation below]

My first attempt at getting this to work again actually looked like this (as addition to cdrom_newpc_intr)

                if (uptodate == 0) {
                        ide_cd_error_cmd(drive, cmd);
                        rq = drive->hwif->rq;
                }

                if (rq) {
                	/* code up to the 2nd complete call */
                }

                if (sense && rc == 2)
                        ide_error(drive, "request sense failure", stat);

[...]

That was before I had any idea why complete was being called twice and that what this is supposed to do won't be done for bio-less requests, anyway, and it worked fine.

Quoted 4 lines

and now here we do the second direct ide_complete_rq and here the rq is freed:

BUG: unable to handle kernel NULL pointer dereference at

0000000000000048

I have just restored the original file to generate another crash dump. [the relevant part of] What I get is EIP == c0251311, edx == 0. This corresponds with the following machine code:

c02512fc <ide_complete_rq>:

c02512fc:       55                      push   %ebp
c02512fd:       89 e5                   mov    %esp,%ebp
c02512ff:       56                      push   %esi
c0251300:       53                      push   %ebx
c0251301:       83 ec 04                sub    $0x4,%esp
c0251304:       89 c3                   mov    %eax,%ebx
c0251306:       89 d0                   mov    %edx,%eax

/* ide_hwif_t *hwif = drive->hwif; */
c0251308: 8b 73 28 mov 0x28(%ebx),%esi

/* struct request *rq = hwif->rq; */
c025130b: 8b 96 c8 01 00 00 mov 0x1c8(%esi),%edx

/* if (blk_noretry_request(rq) .... */

c0251311:       f6 42 24 0e             testb  $0xe,0x24(%edx) /* !!! */
c0251315:       74 04                   je     c025131b <ide_complete_rq+0x1f>

blk_notretry_request is

#define blk_noretry_request(rq)	(blk_failfast_dev(rq) ||	\
				 blk_failfast_transport(rq) ||	\
				 blk_failfast_driver(rq))

and

#define blk_failfast_dev(rq) ((rq)->cmd_flags & REQ_FAILFAST_DEV)
#define blk_failfast_transport(rq) ((rq)->cmd_flags & REQ_FAILFAST_TRANSPORT)
#define blk_failfast_driver(rq) ((rq)->cmd_flags & REQ_FAILFAST_DRIVER)

and

#define REQ_FAILFAST_DEV        (1 << __REQ_FAILFAST_DEV)
#define REQ_FAILFAST_TRANSPORT  (1 << __REQ_FAILFAST_TRANSPORT)
#define REQ_FAILFAST_DRIVER     (1 << __REQ_FAILFAST_DRIVER)

and

enum rq_flag_bits {
__REQ_RW, /* not set, read. set, write */
__REQ_FAILFAST_DEV, /* no driver retries of device errors */
__REQ_FAILFAST_TRANSPORT, /* no driver retries of transport errors */ __REQ_FAILFAST_DRIVER, /* no driver retries of driver errors */
__REQ_DISCARD, /* request to discard sectors */

[...]

This means the values of the relevant __REQ_-constants are 1, 2, and 3 and (1 << 1) | (1 << 2) | (1 << 3) == 2 + 4 + 8 == 14 (== 0xe), hence testb $0xe, 0x24(%edx) (optimized by compiler). edx is 0. 0x24(%edx) is the field at offset 36 in a struct request, which is cmd_flags (on my computer).

blk_end_io always returns zero for bio-less requests. More precisely, it calls end_that_request_data, which is

static int end_that_request_data(struct request *rq, int error,
unsigned int nr_bytes, unsigned int bidi_bytes)

{
        if (rq->bio) {
                if (__end_that_request_first(rq, error, nr_bytes))
                        return 1;

                /* Bidi request must be completed as a whole */
                if (blk_bidi_rq(rq) &&
                    __end_that_request_first(rq->next_rq, error, bidi_bytes))
                        return 1;
        }

        return 0;
}

ie it returns 0 for a request w/o a bio, and looks itself like this:

static int blk_end_io(struct request *rq, int error, unsigned int nr_bytes,
unsigned int bidi_bytes,
int (drv_callback)(struct request *))

{
        struct request_queue *q = rq->q;
        unsigned long flags = 0UL;

        if (end_that_request_data(rq, error, nr_bytes, bidi_bytes))
                return 1;

        /* Special feature for tricky drivers */
        if (drv_callback && drv_callback(rq))
                return 1;

        add_disk_randomness(rq->rq_disk);

        spin_lock_irqsave(q->queue_lock, flags);
        end_that_request_last(rq, error);
        spin_unlock_irqrestore(q->queue_lock, flags);

        return 0;
}

drv_callback is NULL and the 'final return value' is ultimatively returned from the ide_end_rq-call mentioned at the beginning of this overly long e-mail.

An easy way to verify that would be a

	BUG_ON(!rq)

inserted before the blkdev_noretry_request in ide_complete_rq (which I also did -- I have been doing this for long enough to never trust my own assumptions blindly ...).

He who doesn't understand assembly will have a more difficult life because of that :-). While this is interesting, my boss [righfully] hates it and I will now have to do at least an hour of additional overtime.

#7 lore raw

Hi,

Quoted 179 lines · Rainer Weikusat
Quoted 10 lines · Borislav Petkov
Quoted 9 lines · Rainer Weikusat
Quoted 2 lines · Borislav Petkov
Quoted 1 line · Rainer Weikusat

Borislav Petkov <[email protected]> writes:

[...]

Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

so this request is completed as a whole and the rq freed."

Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Technically, this is not quite correct (assuming I haven't overlooked something), because ide_cd_queue_pc still has a reference to the rq.

That doesn't matter because the OOPS happens after the command has been issued and _before_ ide_cd_queue_pc() gets to access the rq ptr again.

Yes. Because the pointer I already mentioned has been reset.

And this happens here (!!!, code from 2.6.30 vanilla):

int ide_complete_rq(ide_drive_t *drive, int error, unsigned int nr_bytes)

{
	ide_hwif_t *hwif = drive->hwif;
	struct request *rq = hwif->rq;
	int rc;

	/*
	 * if failfast is set on a request, override number of sectors
	 * and complete the whole request right now
	 */
	if (blk_noretry_request(rq) && error <= 0)
		nr_bytes = rq->hard_nr_sectors << 9;

	rc = ide_end_rq(drive, rq, error, nr_bytes); 
	if (rc == 0)
		hwif->rq = NULL; /* !!! */

	return rc;
}

[explanation below]

My first attempt at getting this to work again actually looked like this (as addition to cdrom_newpc_intr)

                if (uptodate == 0) {
                        ide_cd_error_cmd(drive, cmd);
                        rq = drive->hwif->rq;
                }

                if (rq) {
                	/* code up to the 2nd complete call */
                }

                if (sense && rc == 2)
                        ide_error(drive, "request sense failure", stat);

[...]

That was before I had any idea why complete was being called twice and that what this is supposed to do won't be done for bio-less requests, anyway, and it worked fine.

Quoted 4 lines

and now here we do the second direct ide_complete_rq and here the rq is freed:

BUG: unable to handle kernel NULL pointer dereference at

0000000000000048

I have just restored the original file to generate another crash dump. [the relevant part of] What I get is EIP == c0251311, edx == 0. This corresponds with the following machine code:

c02512fc <ide_complete_rq>:

c02512fc:       55                      push   %ebp
c02512fd:       89 e5                   mov    %esp,%ebp
c02512ff:       56                      push   %esi
c0251300:       53                      push   %ebx
c0251301:       83 ec 04                sub    $0x4,%esp
c0251304:       89 c3                   mov    %eax,%ebx
c0251306:       89 d0                   mov    %edx,%eax

/* ide_hwif_t *hwif = drive->hwif; */
c0251308: 8b 73 28 mov 0x28(%ebx),%esi

/* struct request *rq = hwif->rq; */
c025130b: 8b 96 c8 01 00 00 mov 0x1c8(%esi),%edx

/* if (blk_noretry_request(rq) .... */

c0251311:       f6 42 24 0e             testb  $0xe,0x24(%edx) /* !!! */
c0251315:       74 04                   je     c025131b <ide_complete_rq+0x1f>

blk_notretry_request is

#define blk_noretry_request(rq)	(blk_failfast_dev(rq) ||	\
				 blk_failfast_transport(rq) ||	\
				 blk_failfast_driver(rq))

and

#define blk_failfast_dev(rq) ((rq)->cmd_flags & REQ_FAILFAST_DEV)
#define blk_failfast_transport(rq) ((rq)->cmd_flags & REQ_FAILFAST_TRANSPORT)
#define blk_failfast_driver(rq) ((rq)->cmd_flags & REQ_FAILFAST_DRIVER)

and

#define REQ_FAILFAST_DEV        (1 << __REQ_FAILFAST_DEV)
#define REQ_FAILFAST_TRANSPORT  (1 << __REQ_FAILFAST_TRANSPORT)
#define REQ_FAILFAST_DRIVER     (1 << __REQ_FAILFAST_DRIVER)

and

enum rq_flag_bits {
__REQ_RW, /* not set, read. set, write */
__REQ_FAILFAST_DEV, /* no driver retries of device errors */
__REQ_FAILFAST_TRANSPORT, /* no driver retries of transport errors */ __REQ_FAILFAST_DRIVER, /* no driver retries of driver errors */
__REQ_DISCARD, /* request to discard sectors */

[...]

This means the values of the relevant __REQ_-constants are 1, 2, and 3 and (1 << 1) | (1 << 2) | (1 << 3) == 2 + 4 + 8 == 14 (== 0xe), hence testb $0xe, 0x24(%edx) (optimized by compiler). edx is 0. 0x24(%edx) is the field at offset 36 in a struct request, which is cmd_flags (on my computer).

blk_end_io always returns zero for bio-less requests. More precisely, it calls end_that_request_data, which is

static int end_that_request_data(struct request *rq, int error,
unsigned int nr_bytes, unsigned int bidi_bytes)

{
        if (rq->bio) {
                if (__end_that_request_first(rq, error, nr_bytes))
                        return 1;

                /* Bidi request must be completed as a whole */
                if (blk_bidi_rq(rq) &&
                    __end_that_request_first(rq->next_rq, error, bidi_bytes))
                        return 1;
        }

        return 0;
}

ie it returns 0 for a request w/o a bio, and looks itself like this:

static int blk_end_io(struct request *rq, int error, unsigned int nr_bytes,
unsigned int bidi_bytes,
int (drv_callback)(struct request *))

{
        struct request_queue *q = rq->q;
        unsigned long flags = 0UL;

        if (end_that_request_data(rq, error, nr_bytes, bidi_bytes))
                return 1;

        /* Special feature for tricky drivers */
        if (drv_callback && drv_callback(rq))
                return 1;

        add_disk_randomness(rq->rq_disk);

        spin_lock_irqsave(q->queue_lock, flags);
        end_that_request_last(rq, error);
        spin_unlock_irqrestore(q->queue_lock, flags);

        return 0;
}

drv_callback is NULL and the 'final return value' is ultimatively returned from the ide_end_rq-call mentioned at the beginning of this overly long e-mail.

An easy way to verify that would be a

	BUG_ON(!rq)
Quoted 7 lines

inserted before the blkdev_noretry_request in ide_complete_rq (which I also did -- I have been doing this for long enough to never trust my own assumptions blindly ...).

He who doesn't understand assembly will have a more difficult life because of that :-).

Well, this is a very good code analysis, I'd say, one for the books :)

Quoted 2 lines

While this is interesting, my boss [righfully] hates it and I will now have to do at least an hour of additional overtime.

Sorry about that, same here :)

Here's another trace I did:

[  266.037646] ide-cd: ide_cd_queue_pc: cmd[0]: 0x51, write: 0x0, timeout: 7000, cmd_flags: 0x8000
[  266.037662] ide-cd: ide_cd_do_request: cmd: 0x51, block: 4294967295, marker: 114
[  266.037667] ide_cd_do_request: dev hdc: type=a, flags=128a440
[  266.037670]   sector 4294967295, nr/cnr 0/0
[  266.037675]   bio dde1e840, biotail dde1e840, buffer (null), len 2
[  266.037678] ide-cd: cdrom_do_block_pc: rq->cmd[0]: 0x51, rq->cmd_type: 0xa
[  266.066559] ide-cd: cdrom_newpc_intr: cmd: 0x51, write: 0x0
[  266.066567] ide-cd: cdrom_newpc_intr: DRQ: stat: 0x58, thislen: 2
[  266.066570] ide-cd: ide_cd_check_ireason: ireason: 0x2, rw: 0x0
[  266.066572] ide-cd: cdrom_newpc_intr: data transfer, rq->cmd_type: 0xa, ireason: 0x2
[  266.149000] ide-cd: cdrom_newpc_intr: cmd: 0x51, write: 0x0
[  266.149010] ide-cd: cdrom_newpc_intr: DRQ: stat: 0x50, thislen: 0
[  266.149014] ide-cd: ide_cd_request_sense_fixup: rq->cmd[0]: 0x51
[  266.149017] ide_complete_rq: completing rq, marker: 114

this is the end of the IRQ handler

[  266.149023] __blk_put_request: marker: 114, ref_count: 2
[  266.149033] ide_cd_queue_pc: putting rq, marker: 114

and here ide_cd_queue_pc comes _after_ that and frees the rq due to refcount == 1.

[  266.149039] __blk_put_request: marker: 114, ref_count: 1
[  266.149045] blk_free_request: marker: 114, ref_count: 0

Now in the fragmented packet command case - that what you call 32/30 - you'll have the first ide_complete_rq() call from ide_cd_error_cmd() and the first decrement of the refcount. Had the rq had a bio, it'd never be come to do a __blk_put_request and decrement the refcount.

But even if it did decrement it, as it does in the real case, that wouldn't be a problem since the rq is still alive (refcount is now 1) and it will be only freed in the second ide_complete_rq() at the end of the IRQ handler.

But, it seems that ide_cd_queue_pc() goes after that first ide_complete_rq() and decrements the refcount, as you suggest, right?

What bothers me here is that we're in IRQ context and running with interrupts disabled so _actually_ the blk_put_request() part of ide_cd_queue_pc() should be getting to run only _after_ the IRQ handler is done and we should be getting the NULL ptr deref in ide_cd_queue_pc(), but we don't. I guess I'm missing something here.

Signature · 2 lines
Regards/Gruss,
    Boris.
#8 lore raw
In reply to Borislav Petkov

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. In order to do this, ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq. The block layer does partial completion only for requests with bio's and if the rq doesn't have one (eg 'GPCMD_READ_DISC_INFO') the request is completed as a whole and the drive->hwif->rq pointer set to NULL afterwards. When calling ide_complete_rq again to report
the error, this null pointer is derefenced, resulting in a kernel crash.

---

ide/ide-cd.c renamed from ide/ide-cd.c.orig +1 -1
@@ -758,7 +758,7 @@ out_end:
 				rq->errors = -EIO;
 		}
 
-		if (uptodate == 0)
+		if (uptodate == 0 && rq->bio)
 			ide_cd_error_cmd(drive, cmd);
 
 		/* make sure it's fully ended */
Tested-by Hans de Bruin
#9 lore raw

Hi,

Quoted 14 lines · Rainer Weikusat

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. In order to do this, ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq. The block layer does partial completion only for requests with bio's and if the rq doesn't have one (eg 'GPCMD_READ_DISC_INFO') the request is completed as a whole and the drive->hwif->rq pointer set to NULL afterwards. When calling ide_complete_rq again to report
the error, this null pointer is derefenced, resulting in a kernel crash.

Sorry, but still not good enough. Instead, I rediffed the change against current linux-ide branch and rewrote the commit message properly keeping your S-O-B.

@Bart: please apply.

--
From: Rainer Weikusat <[email protected]>
Date: Thu, 18 Jun 2009 17:48:24 +0200
Subject: [PATCH] ide-cd: don't do partial completions on bio-less rqs

The block layer completes bio-less requests totally instead of partially and this breaks cdrom drives which fragment packet command data in several DRQ turns (eg GPCMD_READ_DISC_INFO).

More specifically, ide_complete_rq() is called over ide_cd_error_cmd() to partially complete the rq on error. This bio-less request is completed as a whole and when calling ide_complete_rq again to complete the request wholly, the rq is already vanished resulting in the following OOPS:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000048
IP: [<ffffffff804deb50>] ide_complete_rq+0x19/0x4d
PGD 0
Thread overran stack, or stack corrupted
Oops: 0000 [#1] SMP
last sysfs file:
CPU 0
Modules linked in:
Pid: 0, comm: swapper Not tainted 2.6.30-rc8 #22 Precision WorkStation 380
RIP: 0010:[<ffffffff804deb50>]  [<ffffffff804deb50>] ide_complete_rq+0x19/0x4d
RSP: 0018:ffff880028022e18  EFLAGS: 00010096
RAX: 0000000000000002 RBX: ffff88011a84e000 RCX: 0000000000000200
RDX: 0000000000000200 RSI: 0000000000000000 RDI: ffff88011afc9000
RBP: ffff880028022e28 R08: 00000000fffffffb R09: 0000000000000000
R10: ffff8800280302e8 R11: ffff880028030310 R12: 0000000000000000
R13: 0000000000000000 R14: ffff88011afc9000 R15: ffff88011aff7140
FS:  0000000000000000(0000) GS:ffff88002801f000(0000) knlGS:0000000000000000
CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: 0000000000000048 CR3: 0000000000201000 CR4: 00000000000026e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process swapper (pid: 0, threadinfo ffffffff809c0000, task ffffffff808f1360)
Stack:
 ffff88011aecf840 ffff88011aff7140 ffff880028022eb8 ffffffff804ed0ef
 ffffffff8025a811 00ff880028022e70 ffffffff00000050 ffff88011a84e000
 ffff88011a84e108 000000008025ef72 0000000000000000 ffff88011a84e000
Call Trace:
 <IRQ> <0> [<ffffffff804ed0ef>] cdrom_newpc_intr+0x20e/0xac5
 [<ffffffff8025a811>] ? irq_exit+0x47/0x7d
 [<ffffffff804ecee1>] ? cdrom_newpc_intr+0x0/0xac5
 [<ffffffff804de920>] ide_intr+0x1d2/0x220
 [<ffffffff80284ee1>] handle_IRQ_event+0x3a/0xba
 [<ffffffff80286c38>] handle_edge_irq+0xce/0x133
 [<ffffffff8022c0cd>] handle_irq+0x1d/0x29
 [<ffffffff8022b8c7>] do_IRQ+0x5a/0xd5
 [<ffffffff8022a013>] ret_from_intr+0x0/0xa
 <EOI> <0> [<ffffffff80230eee>] ? mwait_idle+0x60/0x6e
 [<ffffffff802282c2>] ? enter_idle+0x20/0x22
 [<ffffffff802286ef>] ? cpu_idle+0x4a/0x8d
 [<ffffffff806ebb05>] ? rest_init+0x65/0x70
 [<ffffffff809cdcef>] ? start_kernel+0x2da/0x3bb
 [<ffffffff809cd271>] ? x86_64_start_reservations+0x81/0xbc
 [<ffffffff809cd37b>] ? x86_64_start_kernel+0xcf/0xf1
Code: c3 48 25 ff ff ff fe 48 89 47 50 e8 a7 86 00 00 eb d7 55 48 89

e5 53 48 83 ec 08 41 89 f0 89 d1 48 8b 5f 40 48 8b b3 28 03 00 00 <f6> 46 48 0e 74 05 45 85 c0 7e 1e 44 89 c2 e8 85 ff ff ff 85 c0 RIP [<ffffffff804deb50>] ide_complete_rq+0x19/0x4d
RSP <ffff880028022e18>
CR2: 0000000000000048

---[ end trace 6662ae44d700bf58 ]---

This fixes http://bugzilla.kernel.org/show_bug.cgi?id=13399.

---
 drivers/ide/ide-cd.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
drivers/ide/ide-cd.c +1 -1
index 0b7645b..4a19686 100644
@@ -667,7 +667,7 @@ out_end:
 				rq->errors = -EIO;
 		}
 
-		if (uptodate == 0)
+		if (uptodate == 0 && rq->bio)
 			ide_cd_error_cmd(drive, cmd);
 
 		/* make sure it's fully ended */
Signature · 7 lines
1.6.3.1



-- 
Regards/Gruss,
Boris
#10 lore raw
Quoted 17 lines · On Thursday 18 June 2009 18:06:34 Borislav Petkov wrote:

Hi,

Quoted 14 lines · Rainer Weikusat

From: Rainer Weikusat <[email protected]>

With 2.6.30, the error handling code in cdrom_newpc_intr was changed to deal with partial request failures by normally completing the 'good' parts of a request and only 'error' the last (and presumably, incompletely transferred) bio associated with a particular request. In order to do this, ide_complete_rq is called over ide_cd_error_cmd() to partially complete the rq. The block layer does partial completion only for requests with bio's and if the rq doesn't have one (eg 'GPCMD_READ_DISC_INFO') the request is completed as a whole and the drive->hwif->rq pointer set to NULL afterwards. When calling ide_complete_rq again to report
the error, this null pointer is derefenced, resulting in a kernel crash.

Rainer, thanks for fixing this bug (with a lot of extra points for the detailed explanation).

Quoted 1 line

@Bart: please apply.

applied [I kept the above patch description]