Skip to contentSkip to conversation

xfs_bmap_add_extent_delay_real: Uninited r[3] corrupts startoff

DISCUSSION Shailendra Tripathi 3 participants 5 messages Oct 9, 2006 – Nov 15, 2006

lore
Share & tools
lore
#0 lore raw

Hi,
It appears that uninitialized r[3] in
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for a particular case.

This sequence is below:

       xfs_bmap_add_extent_delay_real (
       ...
       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
/*
* Filling in the middle part of a previous delayed allocation.
* Contiguity is impossible here.
* This case is avoided almost all the time.
*/
temp = new->br_startoff - PREV.br_startoff; xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff; r[1].br_blockcount = temp2;
xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK); ip->i_df.if_lastex = idx + 1;
ip->i_d.di_nextents++;

Look at extent r[1]. It does not set br_startblock. That is, it is any random value. Now, look at the xfs_bmbt_set_all. Though, it sets the blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
        ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
        r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                 ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                 ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);

Top 21 bits are taken as it is. However, only 9 bit should be taken. So, for random values, it corrupts the startoff which from 9-63 bits.

        r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                 ((xfs_bmbt_rec_base_t)s->br_blockcount &
                 (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being done here. I would appreciate if someone can verify that assertion is correct.

Regards,
Shailendra

#include <stdio.h>
typedef unsigned long __uint64_t;
typedef struct xfs_bmbt_rec_64

{
        __uint64_t              l0, l1;
} xfs_bmbt_rec_64_t;

typedef __uint64_t xfs_bmbt_rec_base_t;
typedef xfs_bmbt_rec_64_t xfs_bmbt_rec_t, xfs_bmdr_rec_t;

typedef enum {
XFS_EXT_NORM, XFS_EXT_UNWRITTEN,
XFS_EXT_DMAPI_OFFLINE
} xfs_exntst_t;

typedef struct xfs_bmbt_irec
{

	__uint64_t   br_startoff;    /* starting file offset */
	__uint64_t   br_startblock;  /* starting block number */
	__uint64_t   br_blockcount;  /* number of blocks */

xfs_exntst_t br_state; /* extent state */
} xfs_bmbt_irec_t;

#define XFS_MASK64LO(n) (((__uint64_t)1 << (n)) - 1)
#define XFS_MASK64HI(n) ((__uint64_t)-1 << (64 - (n)))

int main(void)

{
	xfs_bmbt_irec_t s;
	xfs_bmbt_rec_t  r;
	int extent_flag;

	s.br_startoff = 0;
	s.br_blockcount = 5;
	s.br_startblock = 0xfffffffffffffff0;
	extent_flag = (s.br_state == XFS_EXT_NORM) ? 0 : 1;

	printf("blockcount = 0x%llx\n", s.br_startblock);
	r.l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
	          ((xfs_bmbt_rec_base_t)s.br_startoff << 9) |
	          ((xfs_bmbt_rec_base_t)s.br_startblock >> 43);
	r.l1 = ((xfs_bmbt_rec_base_t)s.br_startblock << 21) |
	         ((xfs_bmbt_rec_base_t)s.br_blockcount &
	         (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

	printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);

	r.l0 = (r.l0 & (xfs_bmbt_rec_base_t)XFS_MASK64HI(55)) |
	        (xfs_bmbt_rec_base_t)((__uint64_t)100 >> 43);
	r.l1 = (r.l1 & (xfs_bmbt_rec_base_t)XFS_MASK64LO(21)) |
	        (xfs_bmbt_rec_base_t)((__uint64_t)100 << 21);

	printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);
	return 0;
}
#1 lore raw

Hi Shailendra,

Quoted 26 lines · Shailendra Tripathi

Hi,
It appears that uninitialized r[3] in
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for a particular case.

This sequence is below:

       xfs_bmap_add_extent_delay_real (
       ...
       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
/*
* Filling in the middle part of a previous delayed allocation.
* Contiguity is impossible here.
* This case is avoided almost all the time.
*/
temp = new->br_startoff - PREV.br_startoff; xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff; r[1].br_blockcount = temp2;
xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK); ip->i_df.if_lastex = idx + 1;
ip->i_d.di_nextents++;

Quoted 8 lines

Look at extent r[1]. It does not set br_startblock. That is, it is any random value. Now, look at the xfs_bmbt_set_all. Though, it sets the blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
        ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
        r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                 ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                 ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);

Top 21 bits are taken as it is. However, only 9 bit should be taken. So, for random values, it corrupts the startoff which from 9-63 bits.

From the code inspection I agree with you that br_startblock doesn't appear to be initialized in this scenario. Otherwise I think the code looks good. If the br_startblock is initialized it should be a value that fits
in 52 bits out of 64 (this is what the ASSERT is for) and the top 12 bits will be 0. The r->l0 gets the top 21 bits of br_startblock, the most significant 12 bits of which are 0 and least significant 9 could be non 0. The r->l1 gets the rest 43 (= 52-9 = 64-21) bits of br_startblock.

I will open a bug report for the uninitialized br_startblock.

Thank you for finding this problem.

Regards,
Vlad

Quoted 66 lines
        r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                 ((xfs_bmbt_rec_base_t)s->br_blockcount &
                 (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being done here. I would appreciate if someone can verify that assertion is correct.

Regards,
Shailendra
------------------------------------------------------------------------

#include <stdio.h>
typedef unsigned long __uint64_t;
typedef struct xfs_bmbt_rec_64

{
        __uint64_t              l0, l1;
} xfs_bmbt_rec_64_t;

typedef __uint64_t xfs_bmbt_rec_base_t; typedef xfs_bmbt_rec_64_t xfs_bmbt_rec_t, xfs_bmdr_rec_t;

typedef enum {
XFS_EXT_NORM, XFS_EXT_UNWRITTEN,
XFS_EXT_DMAPI_OFFLINE
} xfs_exntst_t;

typedef struct xfs_bmbt_irec
{

	__uint64_t   br_startoff;    /* starting file offset */
	__uint64_t   br_startblock;  /* starting block number */
	__uint64_t   br_blockcount;  /* number of blocks */

xfs_exntst_t br_state; /* extent state */
} xfs_bmbt_irec_t;

#define XFS_MASK64LO(n) (((__uint64_t)1 << (n)) - 1)
#define XFS_MASK64HI(n) ((__uint64_t)-1 << (64 - (n)))

int main(void) {
	xfs_bmbt_irec_t s;
	xfs_bmbt_rec_t  r;
	int extent_flag;

	s.br_startoff = 0;
	s.br_blockcount = 5;
	s.br_startblock = 0xfffffffffffffff0;
	extent_flag = (s.br_state == XFS_EXT_NORM) ? 0 : 1;

	printf("blockcount = 0x%llx\n", s.br_startblock);
	r.l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
	          ((xfs_bmbt_rec_base_t)s.br_startoff << 9) |
	          ((xfs_bmbt_rec_base_t)s.br_startblock >> 43);
	r.l1 = ((xfs_bmbt_rec_base_t)s.br_startblock << 21) |
	         ((xfs_bmbt_rec_base_t)s.br_blockcount &
	         (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

	printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);

	r.l0 = (r.l0 & (xfs_bmbt_rec_base_t)XFS_MASK64HI(55)) |
	        (xfs_bmbt_rec_base_t)((__uint64_t)100 >> 43);
	r.l1 = (r.l1 & (xfs_bmbt_rec_base_t)XFS_MASK64LO(21)) |
	        (xfs_bmbt_rec_base_t)((__uint64_t)100 << 21);

	printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);
	return 0;
}
#2 lore raw

This should be all that's needed. This code handles the case where the middle portion of a delayed allocation is being converted and splits the extent into three. The r[1] extent is the rightmost extent that will remain a delayed allocation. Both br_startblock and br_state need to be setup and they will be the same as the original delayed allocation (PREV) so we just inherit those values. Comments?

xfs/xfs_bmap.c 2006-11-02 13:22:41.000000000 +0000 renamed from xfs/xfs_bmap.c_1.358 2006-11-01 14:44:38.000000000 +0000 +1 -0
@@ -1171,6 +1171,7 @@
                xfs_bmap_trace_pre_update(fname, "0", ip, idx, XFS_DATA_FORK);
                xfs_bmbt_set_blockcount(ep, temp);
                r[0] = *new;
+               r[1] = PREV;
                r[1].br_startoff = new_endoff;
                temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff;
                r[1].br_blockcount = temp2;

Lachlan

Quoted 124 lines · Vlad Apostolov

Hi Shailendra,

Quoted 26 lines · Shailendra Tripathi

Hi,
It appears that uninitialized r[3] in
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for a particular case.

This sequence is below:

       xfs_bmap_add_extent_delay_real (
       ...
       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
/*
* Filling in the middle part of a previous delayed allocation.
* Contiguity is impossible here.
* This case is avoided almost all the time.
*/
temp = new->br_startoff - PREV.br_startoff; xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff; r[1].br_blockcount = temp2;
xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK); ip->i_df.if_lastex = idx + 1;
ip->i_d.di_nextents++;

Quoted 8 lines

Look at extent r[1]. It does not set br_startblock. That is, it is any random value. Now, look at the xfs_bmbt_set_all. Though, it sets the blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
        ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
        r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                 ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                 ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);

Top 21 bits are taken as it is. However, only 9 bit should be taken. So, for random values, it corrupts the startoff which from 9-63 bits.

From the code inspection I agree with you that br_startblock doesn't
appear
to be initialized in this scenario. Otherwise I think the code looks good. If the br_startblock is initialized it should be a value that fits
in 52 bits out of 64 (this is what the ASSERT is for) and the top 12 bits will be 0. The r->l0 gets the top 21 bits of br_startblock, the most significant 12 bits of which are 0 and least significant 9 could be non 0. The r->l1 gets the rest 43 (= 52-9 = 64-21) bits of br_startblock.

I will open a bug report for the uninitialized br_startblock.

Thank you for finding this problem.

Regards,
Vlad

Quoted 66 lines
        r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                 ((xfs_bmbt_rec_base_t)s->br_blockcount &
                 (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being done here. I would appreciate if someone can verify that assertion is correct.

Regards,
Shailendra
------------------------------------------------------------------------

#include <stdio.h>
typedef unsigned long __uint64_t;
typedef struct xfs_bmbt_rec_64

{
        __uint64_t              l0, l1;
} xfs_bmbt_rec_64_t;

typedef __uint64_t xfs_bmbt_rec_base_t; typedef xfs_bmbt_rec_64_t xfs_bmbt_rec_t, xfs_bmdr_rec_t;

typedef enum {
XFS_EXT_NORM, XFS_EXT_UNWRITTEN,
XFS_EXT_DMAPI_OFFLINE
} xfs_exntst_t;

typedef struct xfs_bmbt_irec
{

    __uint64_t   br_startoff;    /* starting file offset */
    __uint64_t   br_startblock;  /* starting block number */
    __uint64_t   br_blockcount;  /* number of blocks */

xfs_exntst_t br_state; /* extent state */
} xfs_bmbt_irec_t;

#define XFS_MASK64LO(n) (((__uint64_t)1 << (n)) - 1)
#define XFS_MASK64HI(n) ((__uint64_t)-1 << (64 - (n)))

int main(void) {
    xfs_bmbt_irec_t s;
    xfs_bmbt_rec_t  r;
    int extent_flag;

    s.br_startoff = 0;
    s.br_blockcount = 5;
    s.br_startblock = 0xfffffffffffffff0;
    extent_flag = (s.br_state == XFS_EXT_NORM) ? 0 : 1;

    printf("blockcount = 0x%llx\n", s.br_startblock);
    r.l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
              ((xfs_bmbt_rec_base_t)s.br_startoff << 9) |
              ((xfs_bmbt_rec_base_t)s.br_startblock >> 43);
    r.l1 = ((xfs_bmbt_rec_base_t)s.br_startblock << 21) |
             ((xfs_bmbt_rec_base_t)s.br_blockcount &
             (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);

    r.l0 = (r.l0 & (xfs_bmbt_rec_base_t)XFS_MASK64HI(55)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 >> 43);
    r.l1 = (r.l1 & (xfs_bmbt_rec_base_t)XFS_MASK64LO(21)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 << 21);

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);
    return 0;
}
#3 lore raw

Hi Lachlan,
I would prefer manual assignment here than struct assignment.
r[1].br_startoff and r[1].br_blockcount will be
modified immediately, so it is not worth assigning via ( r[1] = PREV) as it does extra instructions. Compiler would most likely eliminate the extra assignment but, why to leave on the wit of the compiler.

It should be like
r[1].br_state = PREV.br_state;
r[1].br_startblock = 0 ; /* No fancy stuff required here as the aim here is that br_startoff does not get any thing random */

Regards,
Shailendra

Quoted 149 lines · Lachlan McIlroy

This should be all that's needed. This code handles the case where the middle portion of a delayed allocation is being converted and splits the extent into three. The r[1] extent is the rightmost extent that will remain a delayed allocation. Both br_startblock and br_state need to be setup and they will be the same as the original delayed allocation (PREV) so we just inherit those values. Comments?

xfs/xfs_bmap.c 2006-11-02 13:22:41.000000000 +0000 renamed from xfs/xfs_bmap.c_1.358 2006-11-01 14:44:38.000000000 +0000 +1 -0
@@ -1171,6 +1171,7 @@
                xfs_bmap_trace_pre_update(fname, "0", ip, idx, 
 XFS_DATA_FORK);
                xfs_bmbt_set_blockcount(ep, temp);
                r[0] = *new;
+               r[1] = PREV;
                r[1].br_startoff = new_endoff;
                temp2 = PREV.br_startoff + PREV.br_blockcount - 
new_endoff;
                r[1].br_blockcount = temp2;

Lachlan

Quoted 126 lines · Vlad Apostolov

Hi Shailendra,

Quoted 27 lines · Shailendra Tripathi

Hi,
It appears that uninitialized r[3] in
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for a particular case.

This sequence is below:

       xfs_bmap_add_extent_delay_real (
       ...
       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
/*
* Filling in the middle part of a previous delayed
allocation.
* Contiguity is impossible here.
* This case is avoided almost all the time. */
temp = new->br_startoff - PREV.br_startoff; xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff; r[1].br_blockcount = temp2; xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK); ip->i_df.if_lastex = idx + 1; ip->i_d.di_nextents++;

Quoted 8 lines

Look at extent r[1]. It does not set br_startblock. That is, it is any random value. Now, look at the xfs_bmbt_set_all. Though, it sets the blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
        ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
        r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                 ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                 ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);

Top 21 bits are taken as it is. However, only 9 bit should be taken. So, for random values, it corrupts the startoff which from 9-63 bits.

From the code inspection I agree with you that br_startblock doesn't
appear
to be initialized in this scenario. Otherwise I think the code looks good. If the br_startblock is initialized it should be a value that fits
in 52 bits out of 64 (this is what the ASSERT is for) and the top 12 bits will be 0. The r->l0 gets the top 21 bits of br_startblock, the most significant 12 bits of which are 0 and least significant 9 could be non 0. The r->l1 gets the rest 43 (= 52-9 = 64-21) bits of br_startblock.

I will open a bug report for the uninitialized br_startblock.

Thank you for finding this problem.

Regards,
Vlad

Quoted 67 lines
        r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                 ((xfs_bmbt_rec_base_t)s->br_blockcount &
                 (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being done here. I would appreciate if someone can verify that assertion is correct.

Regards,
Shailendra
------------------------------------------------------------------------

#include <stdio.h>
typedef unsigned long __uint64_t;
typedef struct xfs_bmbt_rec_64

{
        __uint64_t              l0, l1;
} xfs_bmbt_rec_64_t;

typedef __uint64_t xfs_bmbt_rec_base_t; typedef xfs_bmbt_rec_64_t xfs_bmbt_rec_t, xfs_bmdr_rec_t;

typedef enum {
XFS_EXT_NORM, XFS_EXT_UNWRITTEN,
XFS_EXT_DMAPI_OFFLINE
} xfs_exntst_t;

typedef struct xfs_bmbt_irec
{

    __uint64_t   br_startoff;    /* starting file offset */
    __uint64_t   br_startblock;  /* starting block number */
    __uint64_t   br_blockcount;  /* number of blocks */

xfs_exntst_t br_state; /* extent state */
} xfs_bmbt_irec_t;

#define XFS_MASK64LO(n) (((__uint64_t)1 << (n)) - 1)
#define XFS_MASK64HI(n) ((__uint64_t)-1 << (64 - (n)))

int main(void) {
    xfs_bmbt_irec_t s;
    xfs_bmbt_rec_t  r;
    int extent_flag;

    s.br_startoff = 0;
    s.br_blockcount = 5;
    s.br_startblock = 0xfffffffffffffff0;
    extent_flag = (s.br_state == XFS_EXT_NORM) ? 0 : 1;

    printf("blockcount = 0x%llx\n", s.br_startblock);
    r.l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
              ((xfs_bmbt_rec_base_t)s.br_startoff << 9) |
              ((xfs_bmbt_rec_base_t)s.br_startblock >> 43);
    r.l1 = ((xfs_bmbt_rec_base_t)s.br_startblock << 21) |
             ((xfs_bmbt_rec_base_t)s.br_blockcount &
             (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);

    r.l0 = (r.l0 & (xfs_bmbt_rec_base_t)XFS_MASK64HI(55)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 >> 43);
    r.l1 = (r.l1 & (xfs_bmbt_rec_base_t)XFS_MASK64LO(21)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 << 21);

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);
    return 0;
}
#4 lore raw

I considered that approach but wasn't keen on setting br_startblock to 0 when it should be NULLSTARTBLOCK. Subsequent calls to xfs_bmbt_set_all() handle NULLSTARTBLOCK differently but the net result ends up being the same and the startblock eventually gets overridden anyway. I'll go with your suggestion.

Quoted 169 lines · Shailendra Tripathi

Hi Lachlan,
I would prefer manual assignment here than struct assignment.
r[1].br_startoff and r[1].br_blockcount will be
modified immediately, so it is not worth assigning via ( r[1] = PREV) as it does extra instructions. Compiler would most likely eliminate the extra assignment but, why to leave on the wit of the compiler.

It should be like
r[1].br_state = PREV.br_state;
r[1].br_startblock = 0 ; /* No fancy stuff required here as the aim here is that br_startoff does not get any thing random */

Regards,
Shailendra

Quoted 152 lines · Lachlan McIlroy

This should be all that's needed. This code handles the case where the middle portion of a delayed allocation is being converted and splits the extent into three. The r[1] extent is the rightmost extent that will remain a delayed allocation. Both br_startblock and br_state need to be setup and they will be the same as the original delayed allocation (PREV) so we just inherit those values. Comments?

xfs/xfs_bmap.c 2006-11-02 13:22:41.000000000 +0000 renamed from xfs/xfs_bmap.c_1.358 2006-11-01 14:44:38.000000000 +0000 +1 -0
@@ -1171,6 +1171,7 @@
                xfs_bmap_trace_pre_update(fname, "0", ip, idx, 
 XFS_DATA_FORK);
                xfs_bmbt_set_blockcount(ep, temp);
                r[0] = *new;
+               r[1] = PREV;
                r[1].br_startoff = new_endoff;
                temp2 = PREV.br_startoff + PREV.br_blockcount - 
new_endoff;
                r[1].br_blockcount = temp2;

Lachlan

Quoted 128 lines · Vlad Apostolov

Hi Shailendra,

Quoted 27 lines · Shailendra Tripathi

Hi,
It appears that uninitialized r[3] in
xfs_bmap_add_extent_delay_real can potentially corrupt the startoff for a particular case.

This sequence is below:

       xfs_bmap_add_extent_delay_real (
       ...
       xfs_bmbt_irec_t         r[3];   /* neighbor extent entries */

case 0:
/*
* Filling in the middle part of a previous delayed
allocation.
* Contiguity is impossible here.
* This case is avoided almost all the time. */
temp = new->br_startoff - PREV.br_startoff; xfs_bmbt_set_blockcount(ep, temp);
r[0] = *new;
r[1].br_startoff = new_endoff;
temp2 = PREV.br_startoff + PREV.br_blockcount - new_endoff; r[1].br_blockcount = temp2; xfs_bmap_insert_exlist(ip, idx + 1, 2, &r[0], XFS_DATA_FORK); ip->i_df.if_lastex = idx + 1; ip->i_d.di_nextents++;

Quoted 8 lines

Look at extent r[1]. It does not set br_startblock. That is, it is any random value. Now, look at the xfs_bmbt_set_all. Though, it sets the blockcount later, the startoff does not get changed.

#if XFS_BIG_BLKNOS
        ASSERT((s->br_startblock & XFS_MASK64HI(12)) == 0);
        r->l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
                 ((xfs_bmbt_rec_base_t)s->br_startoff << 9) |
                 ((xfs_bmbt_rec_base_t)s->br_startblock >> 43);

Top 21 bits are taken as it is. However, only 9 bit should be taken. So, for random values, it corrupts the startoff which from 9-63 bits.

From the code inspection I agree with you that br_startblock doesn't
appear
to be initialized in this scenario. Otherwise I think the code looks good. If the br_startblock is initialized it should be a value that fits
in 52 bits out of 64 (this is what the ASSERT is for) and the top 12 bits will be 0. The r->l0 gets the top 21 bits of br_startblock, the most significant 12 bits of which are 0 and least significant 9 could be non 0. The r->l1 gets the rest 43 (= 52-9 = 64-21) bits of br_startblock.

I will open a bug report for the uninitialized br_startblock.

Thank you for finding this problem.

Regards,
Vlad

Quoted 67 lines
        r->l1 = ((xfs_bmbt_rec_base_t)s->br_startblock << 21) |
                 ((xfs_bmbt_rec_base_t)s->br_blockcount &
                 (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

I have attached a small program which does the same thing as it is being done here. I would appreciate if someone can verify that assertion is correct.

Regards,
Shailendra
------------------------------------------------------------------------

#include <stdio.h>
typedef unsigned long __uint64_t;
typedef struct xfs_bmbt_rec_64

{
        __uint64_t              l0, l1;
} xfs_bmbt_rec_64_t;

typedef __uint64_t xfs_bmbt_rec_base_t; typedef xfs_bmbt_rec_64_t xfs_bmbt_rec_t, xfs_bmdr_rec_t;

typedef enum {
XFS_EXT_NORM, XFS_EXT_UNWRITTEN,
XFS_EXT_DMAPI_OFFLINE
} xfs_exntst_t;

typedef struct xfs_bmbt_irec
{

    __uint64_t   br_startoff;    /* starting file offset */
    __uint64_t   br_startblock;  /* starting block number */
    __uint64_t   br_blockcount;  /* number of blocks */

xfs_exntst_t br_state; /* extent state */
} xfs_bmbt_irec_t;

#define XFS_MASK64LO(n) (((__uint64_t)1 << (n)) - 1)
#define XFS_MASK64HI(n) ((__uint64_t)-1 << (64 - (n)))

int main(void) {
    xfs_bmbt_irec_t s;
    xfs_bmbt_rec_t  r;
    int extent_flag;

    s.br_startoff = 0;
    s.br_blockcount = 5;
    s.br_startblock = 0xfffffffffffffff0;
    extent_flag = (s.br_state == XFS_EXT_NORM) ? 0 : 1;

    printf("blockcount = 0x%llx\n", s.br_startblock);
    r.l0 = ((xfs_bmbt_rec_base_t)extent_flag << 63) |
              ((xfs_bmbt_rec_base_t)s.br_startoff << 9) |
              ((xfs_bmbt_rec_base_t)s.br_startblock >> 43);
    r.l1 = ((xfs_bmbt_rec_base_t)s.br_startblock << 21) |
             ((xfs_bmbt_rec_base_t)s.br_blockcount &
             (xfs_bmbt_rec_base_t)XFS_MASK64LO(21));

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);

    r.l0 = (r.l0 & (xfs_bmbt_rec_base_t)XFS_MASK64HI(55)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 >> 43);
    r.l1 = (r.l1 & (xfs_bmbt_rec_base_t)XFS_MASK64LO(21)) |
            (xfs_bmbt_rec_base_t)((__uint64_t)100 << 21);

    printf("l0 = 0x%llx l1 = 0x%llx\n", r.l0, r.l1);
    return 0;
}