IDE crash...
I'm debugging a blk_rq_map_sg() crash that i'm getting on sparc64 as root is mounted over IDE. I think I know what is happening now.
The IDE sg table is allocated and initialized like this in drivers/ide/ide-probe.c:
x = kmalloc(sizeof(struct scatterlist) * nents, GFP_XXX); sg_init_table(x, nents);
So far, so good.
Now, ide_map_sg() passes requests down to blk_rq_map_sg() like this in drivers/block/ide-io.c:
hwif->sg_nents = blk_rq_map_sg(drive->queue, rq, sg);
Ok, so what does blk_rq_map_sg() do?
sg = NULL;
rq_for_each_segment(bvec, rq, iter) {
...
if (bvprv && cluster) {
...
} else {
new_segment:
if (!sg)
sg = sglist;
else
sg = sg_next(sg);
...
}
bvprv = bvec;
} /* segments in rq */
if (sg)
__sg_mark_end(sg);So let's say the first request comes in and needs 2 segs. This will mark sg[1].page_link with 0x2
If the next request from IDE needs 4 segs, we'll OOPS because sg_next() on &sg[1] will see page_link bit 0x2 is set and therefore return NULL.
A quick look shows that if you're testing on SCSI (or something layered on top of it like SATA or PATA) you won't see this seemingly guarenteed crash because the SCSI mid-layer allocates a fresh sglist via mempool_alloc() and runs sg_init_table() on it for every I/O request.
Quoted 45 lines · David Miller
I'm debugging a
blk_rq_map_sg()crash that i'm getting on sparc64 as root is mounted over IDE. I think I know what is happening now.The IDE sg table is allocated and initialized like this in
drivers/ide/ide-probe.c:x = kmalloc(sizeof(struct scatterlist) * nents, GFP_XXX); sg_init_table(x, nents);So far, so good.
Now,
ide_map_sg()passes requests down toblk_rq_map_sg()like this indrivers/block/ide-io.c:hwif->sg_nents = blk_rq_map_sg(drive->queue, rq, sg);Ok, so what does
blk_rq_map_sg()do?sg = NULL; rq_for_each_segment(bvec, rq, iter) { ... if (bvprv && cluster) { ... } else { new_segment: if (!sg) sg = sglist; else sg = sg_next(sg); ... } bvprv = bvec; } /* segments in rq */ if (sg) __sg_mark_end(sg);So let's say the first request comes in and needs 2 segs. This will mark sg[1].page_link with
0x2If the next request from IDE needs 4 segs, we'll OOPS because
sg_next()on &sg[1] will seepage_linkbit0x2is set and therefore return NULL.
Quoted 5 lines
A quick look shows that if you're testing on SCSI (or something layered on top of it like SATA or PATA) you won't see this seemingly guarenteed crash because the SCSI mid-layer allocates a fresh sglist via
mempool_alloc()and runssg_init_table()on it for every I/O request.
We should never see the end pointer in blk_rq_map_sg(), or that's a bug in the driver. So it should be OK to just clear the end pointer always in there, even if it's not the prettiest solution...
This just needs to be wrapped up in some scatterlist.h macro/function.
block/ll_rw_blk.c +6 -0
@@ -1354,6 +1354,12 @@ new_segment:else sg =sg_next (sg); +/* +* Clear end-of-table pointer, we'll mark a new one +* at the end +*/ + sg->page_link &= ~0x2 ; +sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;sg_set_page (sg, bvec->bv_page);
Signature · 1 line
Jens Axboe
Quoted 65 lines · Jens Axboe
Quoted 45 lines · David Miller
I'm debugging a
blk_rq_map_sg()crash that i'm getting on sparc64 as root is mounted over IDE. I think I know what is happening now.The IDE sg table is allocated and initialized like this in
drivers/ide/ide-probe.c:x = kmalloc(sizeof(struct scatterlist) * nents, GFP_XXX); sg_init_table(x, nents);So far, so good.
Now,
ide_map_sg()passes requests down toblk_rq_map_sg()like this indrivers/block/ide-io.c:hwif->sg_nents = blk_rq_map_sg(drive->queue, rq, sg);Ok, so what does
blk_rq_map_sg()do?sg = NULL; rq_for_each_segment(bvec, rq, iter) { ... if (bvprv && cluster) { ... } else { new_segment: if (!sg) sg = sglist; else sg = sg_next(sg); ... } bvprv = bvec; } /* segments in rq */ if (sg) __sg_mark_end(sg);So let's say the first request comes in and needs 2 segs. This will mark sg[1].page_link with
0x2If the next request from IDE needs 4 segs, we'll OOPS because
sg_next()on &sg[1] will seepage_linkbit0x2is set and therefore return NULL.Quoted 5 lines
A quick look shows that if you're testing on SCSI (or something layered on top of it like SATA or PATA) you won't see this seemingly guarenteed crash because the SCSI mid-layer allocates a fresh sglist via
mempool_alloc()and runssg_init_table()on it for every I/O request.We should never see the end pointer in
blk_rq_map_sg(), or that's a bug in the driver. So it should be OK to just clear the end pointer always in there, even if it's not the prettiest solution...This just needs to be wrapped up in some
scatterlist.hmacro/function.block/ll_rw_blk.c +0
-0@@ -1354,6 +1354,12 @@ new_segment:else sg =sg_next (sg);
block/ll_rw_blk.c new_segment: · line 1365
Quoted 10 lines
block/ll_rw_blk.c +6
-0+/* +* Clear end-of-table pointer, we'll mark a new one +* at the end +*/ + sg->page_link &= ~0x2 ; +sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;sg_set_page (sg, bvec->bv_page);
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +1 -0
@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );
drivers/ide/ide-probe.c +0 -2
@@ -1324,8 +1324,6 @@ static int hwif_init(ide_hwif_t *hwif) goto out; } -sg_init_table (hwif->sg_table, hwif->sg_max_nents); -if (init_irq (hwif) ==0 ) goto done;
Signature · 1 line
Jens Axboe
Quoted 67 lines · On Tue, 23 Oct 2007 09:09:33 +0200 Jens Axboe <[email protected]> wrote:
Quoted 66 lines · Jens Axboe
Quoted 45 lines · David Miller
I'm debugging a
blk_rq_map_sg()crash that i'm getting on sparc64 as root is mounted over IDE. I think I know what is happening now.The IDE sg table is allocated and initialized like this in
drivers/ide/ide-probe.c:x = kmalloc(sizeof(struct scatterlist) * nents, GFP_XXX); sg_init_table(x, nents);So far, so good.
Now,
ide_map_sg()passes requests down toblk_rq_map_sg()like this indrivers/block/ide-io.c:hwif->sg_nents = blk_rq_map_sg(drive->queue, rq, sg);Ok, so what does
blk_rq_map_sg()do?sg = NULL; rq_for_each_segment(bvec, rq, iter) { ... if (bvprv && cluster) { ... } else { new_segment: if (!sg) sg = sglist; else sg = sg_next(sg); ... } bvprv = bvec; } /* segments in rq */ if (sg) __sg_mark_end(sg);So let's say the first request comes in and needs 2 segs. This will mark sg[1].page_link with
0x2If the next request from IDE needs 4 segs, we'll OOPS because
sg_next()on &sg[1] will seepage_linkbit0x2is set and therefore return NULL.Quoted 5 lines
A quick look shows that if you're testing on SCSI (or something layered on top of it like SATA or PATA) you won't see this seemingly guarenteed crash because the SCSI mid-layer allocates a fresh sglist via
mempool_alloc()and runssg_init_table()on it for every I/O request.We should never see the end pointer in
blk_rq_map_sg(), or that's a bug in the driver. So it should be OK to just clear the end pointer always in there, even if it's not the prettiest solution...This just needs to be wrapped up in some
scatterlist.hmacro/function.block/ll_rw_blk.c +0
-0@@ -1354,6 +1354,12 @@ new_segment:else sg =sg_next (sg);
Quoted 12 lines
block/ll_rw_blk.c new_segment: · line 1365
Quoted 9 lines
block/ll_rw_blk.c +6
-0+/* +* Clear end-of-table pointer, we'll mark a new one +* at the end +*/ + sg->page_link &= ~0x2 ; +sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;sg_set_page (sg, bvec->bv_page);Eh this wont work, it's the wrong entry... Here's a temporary work-around.
Yeah, it won't work. Now we must call sg_init_table for every I/O request (it's not nice).
I think that there are other blk_rq_map_sg users need such fix.
Quoted 29 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );drivers/ide/ide-probe.c +0
-2@@ -1324,8 +1324,6 @@ static int hwif_init(ide_hwif_t *hwif) goto out; } -sg_init_table (hwif->sg_table, hwif->sg_max_nents); -if (init_irq (hwif) ==0 ) goto done;Signature · 2 lines
-- Jens Axboe
Quoted 83 lines · FUJITA Tomonori
Quoted 67 lines · On Tue, 23 Oct 2007 09:09:33 +0200 Jens Axboe <[email protected]> wrote:
Quoted 66 lines · Jens Axboe
Quoted 45 lines · David Miller
I'm debugging a
blk_rq_map_sg()crash that i'm getting on sparc64 as root is mounted over IDE. I think I know what is happening now.The IDE sg table is allocated and initialized like this in
drivers/ide/ide-probe.c:x = kmalloc(sizeof(struct scatterlist) * nents, GFP_XXX); sg_init_table(x, nents);So far, so good.
Now,
ide_map_sg()passes requests down toblk_rq_map_sg()like this indrivers/block/ide-io.c:hwif->sg_nents = blk_rq_map_sg(drive->queue, rq, sg);Ok, so what does
blk_rq_map_sg()do?sg = NULL; rq_for_each_segment(bvec, rq, iter) { ... if (bvprv && cluster) { ... } else { new_segment: if (!sg) sg = sglist; else sg = sg_next(sg); ... } bvprv = bvec; } /* segments in rq */ if (sg) __sg_mark_end(sg);So let's say the first request comes in and needs 2 segs. This will mark sg[1].page_link with
0x2If the next request from IDE needs 4 segs, we'll OOPS because
sg_next()on &sg[1] will seepage_linkbit0x2is set and therefore return NULL.Quoted 5 lines
A quick look shows that if you're testing on SCSI (or something layered on top of it like SATA or PATA) you won't see this seemingly guarenteed crash because the SCSI mid-layer allocates a fresh sglist via
mempool_alloc()and runssg_init_table()on it for every I/O request.We should never see the end pointer in
blk_rq_map_sg(), or that's a bug in the driver. So it should be OK to just clear the end pointer always in there, even if it's not the prettiest solution...This just needs to be wrapped up in some
scatterlist.hmacro/function.block/ll_rw_blk.c +0
-0@@ -1354,6 +1354,12 @@ new_segment:else sg =sg_next (sg);Quoted 12 lines
block/ll_rw_blk.c new_segment: · line 1365
Quoted 9 lines
block/ll_rw_blk.c +6
-0+/* +* Clear end-of-table pointer, we'll mark a new one +* at the end +*/ + sg->page_link &= ~0x2 ; +sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;sg_set_page (sg, bvec->bv_page);Eh this wont work, it's the wrong entry... Here's a temporary work-around.
Quoted 2 lines
Yeah, it won't work. Now we must call
sg_init_tablefor every I/O request (it's not nice).
I think the fix would be to have a sg_next_and_clear() or something that doesn't honor the 0x02 termination bit and clears it, for the cases where you KNOW that there are more entries.
Quoted 1 line
I think that there are other
blk_rq_map_sgusers need such fix.
Possibly, I did do quite a few of them. Alternatively, we can remove __sg_mark_end() and leave that up to the driver.
block/ll_rw_blk.c +1 -1
@@ -1352,7 +1352,7 @@ new_segment:if (!sg) sg = sglist;else - sg =sg_next (sg); + sg =sg_next_force (sg);sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;
include/linux/scatterlist.h +16 -0
@@ -99,6 +99,22 @@ static inline struct scatterlist *sg_next(struct scatterlist *sg)return sg; } +/** +* sg_next_force - return the next scatterlist entry in a list +* @sg: The current sg entry +* +* Description: +* Must only be used when more entries beyond this one is known to exist, +* as it clears the termination bit. Useful to avoid adding a full sg +* table init on every mapping. +* +**/ +static inline struct scatterlist *sg_next_force (struct scatterlist *sg) +{ + sg->page_link &= ~0x02 ; +return sg_next (sg); +} +/* * Loop over each sg element, following the pointer to a new list if necessary */
Signature · 1 line
Jens Axboe
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200
Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0
drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );
That's the exact patch I'm about to boot test :-)
Quoted 13 lines · David Miller
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 10 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;
Quoted 7 lines
drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 5 lines
drivers/ide/ide-io.c +1
-0if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
That should work - once you verify that, would you mind testing this one as well? Thanks!
block/ll_rw_blk.c +1 -1
@@ -1352,7 +1352,7 @@ new_segment:if (!sg) sg = sglist;else - sg =sg_next (sg); + sg =sg_next_force (sg);sg_dma_len (sg) =0 ;sg_dma_address (sg) =0 ;
include/linux/scatterlist.h +16 -0
@@ -99,6 +99,22 @@ static inline struct scatterlist *sg_next(struct scatterlist *sg)return sg; } +/** +* sg_next_force - return the next scatterlist entry in a list +* @sg: The current sg entry +* +* Description: +* Must only be used when more entries beyond this one is known to exist, +* as it clears the termination bit. Useful to avoid adding a full sg +* table init on every mapping. +* +**/ +static inline struct scatterlist *sg_next_force (struct scatterlist *sg) +{ + sg->page_link &= ~0x02 ; +return sg_next (sg); +} +/* * Loop over each sg element, following the pointer to a new list if necessary */
Signature · 1 line
Jens Axboe
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200
Quoted 22 lines · David Miller
Quoted 13 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 10 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;Quoted 7 lines
drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 5 lines
drivers/ide/ide-io.c +1
-0if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Quoted 25 lines · David Miller
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Great, thanks for testing that as well. Thinking a bit more about it, I think the forced clear should stay in blk_rq_map_sg() since we don't want to advertise it to drivers. So I'll just open-code it in there.
Signature · 1 line
Jens Axboe
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 25 lines · David Miller
Quoted 25 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Jens> Great, thanks for testing that as well. Thinking a bit more Jens> about it, I think the forced clear should stay in Jens> blk_rq_map_sg() since we don't want to advertise it to Jens> drivers. So I'll just open-code it in there.
Should there be more bounds checking here, so that if you try to do a _force() you at least check to make sure that there's something beyond there and useable on the list?
We're saving the time from not reallocating from scratch, but let's make it robust by doing at least some more checks here.
John
Quoted 38 lines · John Stoffel
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 25 lines · David Miller
Quoted 25 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Jens> Great, thanks for testing that as well. Thinking a bit more Jens> about it, I think the forced clear should stay in Jens>
blk_rq_map_sg()since we don't want to advertise it to Jens> drivers. So I'll just open-code it in there.
Quoted 6 lines
Should there be more bounds checking here, so that if you try to do a
_force()you at least check to make sure that there's something beyond there and useable on the list?We're saving the time from not reallocating from scratch, but let's make it robust by doing at least some more checks here.
We have to rely on the caller passing in an sgtable that is big enough to map the request, we have always made that assumption. Anything else would be a bug, of course. Now, catching that bug would indeed be nice. Any suggestions?
Signature · 1 line
Jens Axboe
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 2 lines · John Stoffel
Quoted 2 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 25 lines · David Miller
Quoted 25 lines
Quoted 25 lines
Quoted 25 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Jens> Great, thanks for testing that as well. Thinking a bit more Jens> about it, I think the forced clear should stay in Jens> blk_rq_map_sg() since we don't want to advertise it to Jens> drivers. So I'll just open-code it in there.
Quoted 7 lines
Quoted 7 lines
Should there be more bounds checking here, so that if you try to do a
_force()you at least check to make sure that there's something beyond there and useable on the list?We're saving the time from not reallocating from scratch, but let's make it robust by doing at least some more checks here.
Jens> We have to rely on the caller passing in an sgtable that is big enough Jens> to map the request, we have always made that assumption. Anything else Jens> would be a bug, of course. Now, catching that bug would indeed be nice. Jens> Any suggestions?
Poor mans slab poisoning maybe? As Alan Cox was saying about keeping it simple with a NULL entry on the end of the SG list, having an easy way to not fall of the end seems key.
But from reading and re-reading the thread and some of the code, it looks like this list is really allocated, then the driver can just re-use the list (or a subset) as much as it likes without clearing and then reallocating.
So if you've got a chain of 127 items, does it make sense when you only use 16 of them to poison the next three or four in the chain to really make sure you don't fall off the end?
Sorry I don't have a concrete example, my C-foo is quite weak and I haven't had the time to really dive into this code. So take my comments as just an interested but ignorant bystander butting in. :]
John
Quoted 57 lines · John Stoffel
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 2 lines · John Stoffel
Quoted 2 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
"Jens" == Jens Axboe <[email protected]> writes:
Quoted 25 lines · David Miller
Quoted 25 lines
Quoted 25 lines
Quoted 25 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
Jens> Great, thanks for testing that as well. Thinking a bit more Jens> about it, I think the forced clear should stay in Jens>
blk_rq_map_sg()since we don't want to advertise it to Jens> drivers. So I'll just open-code it in there.Quoted 7 lines
Quoted 7 lines
Should there be more bounds checking here, so that if you try to do a
_force()you at least check to make sure that there's something beyond there and useable on the list?We're saving the time from not reallocating from scratch, but let's make it robust by doing at least some more checks here.
Jens> We have to rely on the caller passing in an sgtable that is big enough Jens> to map the request, we have always made that assumption. Anything else Jens> would be a bug, of course. Now, catching that bug would indeed be nice. Jens> Any suggestions?
Poor mans slab poisoning maybe? As Alan Cox was saying about keeping it simple with a NULL entry on the end of the SG list, having an easy way to not fall of the end seems key.
Quoted 8 lines
But from reading and re-reading the thread and some of the code, it looks like this list is really allocated, then the driver can just re-use the list (or a subset) as much as it likes without clearing and then reallocating.
So if you've got a chain of 127 items, does it make sense when you only use 16 of them to poison the next three or four in the chain to really make sure you don't fall off the end?
Yes, that is indeed how most drivers work. It's both cheaper and faster to keep the list than allocating and initializing it every time. With the current termination bit approach, only that bits needs to be set/cleared to properly terminate the sgtable. So it's free, since it happens with setting page/length/offset.
Signature · 1 line
Jens Axboe
Quoted 25 lines · On Tue, 23 Oct 2007 00:43:21 -0700 (PDT) David Miller <[email protected]>…
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
BTW, we avoid calling sg_init_table() for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr in blk_rq_map_sg(). So sparc64 iommu needs to do that for itself?
Quoted 31 lines · FUJITA Tomonori
Quoted 28 lines · On Tue, 23 Oct 2007 00:43:21 -0700 (PDT) David Miller <[email protected]>…
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 24 lines · David Miller
Quoted 13 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 10 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;Quoted 7 lines
drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 5 lines
drivers/ide/ide-io.c +1
-0if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
That should work - once you verify that, would you mind testing this one as well? Thanks!
Quoted 7 lines
Quoted 1 line
This one works here too, thanks.
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?
It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
Signature · 1 line
Jens Axboe
Quoted 38 lines · Jens Axboe
Quoted 37 lines · FUJITA Tomonori
Quoted 25 lines · On Tue, 23 Oct 2007 00:43:21 -0700 (PDT) David Miller <[email protected]>…
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?
Quoted 5 lines
Quoted 1 line
It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
BTW, you sure that should not include to check whether sg_next() returns non-NULL?
Signature · 1 line
Jens Axboe
Quoted 39 lines · On Tue, 23 Oct 2007 12:58:11 +0200 Jens Axboe <[email protected]> wrote:
Quoted 38 lines · Jens Axboe
Quoted 37 lines · FUJITA Tomonori
Quoted 25 lines · On Tue, 23 Oct 2007 00:43:21 -0700 (PDT) David Miller <[email protected]>…
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
BTW, you sure that should not include to check whether
sg_next()returns non-NULL?
Yeah, the following part checks that:
+ if (dma_sg != sg) {
Quoted 53 lines · FUJITA Tomonori
Quoted 39 lines · On Tue, 23 Oct 2007 12:58:11 +0200 Jens Axboe <[email protected]> wrote:
Quoted 38 lines · Jens Axboe
Quoted 37 lines · FUJITA Tomonori
Quoted 25 lines · On Tue, 23 Oct 2007 00:43:21 -0700 (PDT) David Miller <[email protected]>…
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:23:59 +0200Quoted 22 lines · David Miller
Quoted 20 lines
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 09:09:33 +0200Quoted 8 lines
Eh this wont work, it's the wrong entry... Here's a temporary work-around.
drivers/ide/ide-io.c +0
-0drivers/ide/ide-io.c void ide_map_sg(ide_drive_t *drive, struct request *rq) · line 828
Quoted 7 lines
drivers/ide/ide-io.c +1
-0@@ -822,6 +822,7 @@ void ide_map_sg(ide_drive_t *drive, struct request *rq)return ;if (rq->cmd_type !=REQ_TYPE_ATA_TASKFILE ) { +sg_init_table (hwif->sg_table, hwif->sg_max_nents); hwif->sg_nents =blk_rq_map_sg (drive->queue, rq, sg); }else {sg_init_one (sg, rq->buffer, rq->nr_sectors *SECTOR_SIZE );That's the exact patch I'm about to boot test :-)
Quoted 4 lines
Quoted 2 lines
That should work - once you verify that, would you mind testing this one as well? Thanks!
This one works here too, thanks.
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?Quoted 8 lines
Quoted 5 lines
Quoted 1 line
It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
BTW, you sure that should not include to check whether
sg_next()returns non-NULL?Yeah, the following part checks that:
Review
Quoted 1 line
(quoted diff) +1
-0+ if (dma_sg != sg) {
Ah good, thanks for confirming.
Signature · 1 line
Jens Axboe
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 12:57:11 +0200
Quoted 10 lines · FUJITA Tomonori
Quoted 5 lines
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
No objections.
Quoted 8 lines · David Miller
From: Jens Axboe <[email protected]>
Date: Tue, 23 Oct 2007 12:57:11 +0200Quoted 5 lines · FUJITA Tomonori
Quoted 4 lines
BTW, we avoid calling
sg_init_table()for every I/O request due to Jens' trick. And Jens will remove the code to clear sg_dma_len/addr inblk_rq_map_sg(). So sparc64 iommu needs to do that for itself?
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
It does - Dave, do you agree with the patch? I think it should be added, so that sparc64 isn't the odd-arch-out in this respect. I've tentatively applied the patch, Davem?
No objections.
Great, thanks Dave.
Signature · 1 line
Jens Axboe