scsi: target: tcmu: Replace IDR and radix_tree with XArray
Subsystems (3)
Bodo Stroesser
maintainercover scsi: target: tcmu: Replace IDR and radix_tree with XArray
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed - radix_tree_for_each_contig macro.
Since general direction is to use xarray as replacement for radix_tree and IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog to radix_tree_for_each_contig.
Bodo Stroesser (2):
scsi: target: tcmu: Replace IDR by XArray
scsi: target: tcmu: Replace radix_tree with XArray
drivers/target/target_core_user.c | 64 +++++++++++++++++++-------------------- 1 file changed, 31 insertions(+), 33 deletions(-)
Signature · 1 line
2.12.3
Bodo Stroesser
maintainerPATCH 1/2 scsi: target: tcmu: Replace IDR by XArray
An attempt from Matthew Wilcox to replace IDR usage by XArray in tcmu more than 1 year ago unfortunately got lost.
I rebased that work on latest tcmu and tested it.
--- drivers/target/target_core_user.c | 30 +++++++++++++++--------------- 1 file changed, 15 insertions(+), 15 deletions(-)
drivers/target/target_core_user.c +15 -15
@@ -8,12 +8,12 @@#include <linux/spinlock.h> #include <linux/module.h> -#include <linux/idr.h> #include <linux/kernel.h> #include <linux/timer.h> #include <linux/parser.h> #include <linux/vmalloc.h> #include <linux/uio_driver.h> +#include <linux/xarray.h> #include <linux/radix-tree.h> #include <linux/stringify.h> #include <linux/bitops.h>
@@ -145,7 +145,7 @@ struct tcmu_dev {unsigned long *data_bitmap;struct radix_tree_root data_blocks; -struct idr commands; +struct xarray commands;struct timer_list cmd_timer;unsigned int cmd_time_out;
@@ -977,8 +977,8 @@ static int queue_cmd_ring(struct tcmu_cmd *tcmu_cmd, sense_reason_t *scsi_err)struct tcmu_mailbox *mb = udev->mb_addr;struct tcmu_cmd_entry *entry;struct iovec *iov; -int iov_cnt, iov_bidi_cnt, cmd_id; -uint32_t cmd_head; +int iov_cnt, iov_bidi_cnt; +uint32_t cmd_id, cmd_head;uint64_t cdb_off;/* size of data buffer needed */ size_t data_length = (size_t )tcmu_cmd->dbi_cnt *DATA_BLOCK_SIZE ;
@@ -1031,8 +1031,8 @@ static int queue_cmd_ring(struct tcmu_cmd *tcmu_cmd, sense_reason_t *scsi_err)*/ goto free_and_queue; - cmd_id =idr_alloc (&udev->commands, tcmu_cmd,1 ,USHRT_MAX ,GFP_NOWAIT ); -if (cmd_id <0 ) { +if (xa_alloc (&udev->commands, &cmd_id, tcmu_cmd,XA_LIMIT (1 ,0xffff ), +GFP_NOWAIT ) <0 ) {pr_err ("tcmu: Could not allocate cmd id.\n" );tcmu_cmd_free_data (tcmu_cmd, tcmu_cmd->dbi_cnt);
@@ -1415,7 +1415,7 @@ static unsigned int tcmu_handle_completions(struct tcmu_dev *udev) }WARN_ON (tcmu_hdr_get_op (entry->hdr.len_op) !=TCMU_OP_CMD ); - cmd =idr_remove (&udev->commands, entry->hdr.cmd_id); + cmd =xa_erase (&udev->commands, entry->hdr.cmd_id);if (!cmd) {pr_err ("cmd_id %u not found, ring is broken\n" , entry->hdr.cmd_id);
@@ -1433,7 +1433,7 @@ static unsigned int tcmu_handle_completions(struct tcmu_dev *udev) free_space =tcmu_run_tmr_queue (udev);if (atomic_read (&global_db_count) > tcmu_global_max_blocks && -idr_is_empty (&udev->commands) &&list_empty (&udev->qfull_queue)) { +xa_empty (&udev->commands) &&list_empty (&udev->qfull_queue)) {/* * Allocated blocks exceeded global block limit, currently no * more pending or waiting commands so try to reclaim blocks.
@@ -1556,7 +1556,7 @@ static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)INIT_LIST_HEAD (&udev->qfull_queue);INIT_LIST_HEAD (&udev->tmr_queue);INIT_LIST_HEAD (&udev->inflight_queue); -idr_init (&udev->commands); +xa_init_flags (&udev->commands,XA_FLAGS_ALLOC1 );timer_setup (&udev->qfull_timer, tcmu_qfull_timedout,0 );timer_setup (&udev->cmd_timer, tcmu_cmd_timedout,0 );
@@ -1616,7 +1616,7 @@ static void tcmu_dev_kref_release(struct kref *kref)struct se_device *dev = &udev->se_dev;struct tcmu_cmd *cmd;bool all_expired = true; -int i; +unsigned long i;vfree (udev->mb_addr); udev->mb_addr =NULL ;
@@ -1628,7 +1628,7 @@ static void tcmu_dev_kref_release(struct kref *kref)/* Upper layer should drain all requests before calling this */ mutex_lock (&udev->cmdr_lock); -idr_for_each_entry (&udev->commands, cmd, i) { +xa_for_each (&udev->commands, i, cmd) {if (tcmu_check_and_free_pending_cmd (cmd) !=0 ) all_expired = false; }
@@ -1636,7 +1636,7 @@ static void tcmu_dev_kref_release(struct kref *kref)tcmu_remove_all_queued_tmr (udev);if (!list_empty (&udev->qfull_queue)) all_expired = false; -idr_destroy (&udev->commands); +xa_destroy (&udev->commands);WARN_ON (!all_expired);tcmu_blocks_release (&udev->data_blocks,0 , udev->dbi_max +1 );
@@ -2226,16 +2226,16 @@ static void tcmu_reset_ring(struct tcmu_dev *udev, u8 err_level) {struct tcmu_mailbox *mb;struct tcmu_cmd *cmd; -int i; +unsigned long i;mutex_lock (&udev->cmdr_lock); -idr_for_each_entry (&udev->commands, cmd, i) { +xa_for_each (&udev->commands, i, cmd) {pr_debug ("removing cmd %u on dev %s from ring (is expired %d)\n" , cmd->cmd_id, udev->name,test_bit (TCMU_CMD_BIT_EXPIRED , &cmd->flags)); -idr_remove (&udev->commands, i); +xa_erase (&udev->commands, i);if (!test_bit (TCMU_CMD_BIT_EXPIRED , &cmd->flags)) {WARN_ON (!cmd->se_cmd);list_del_init (&cmd->queue_entry);
Signature · 1 line
2.12.3
Bodo Stroesser
maintainerPATCH 2/2 scsi: target: tcmu: Replace radix_tree with XArray
In reply to Bodo StroesserAn attempt from Matthew Wilcox to replace radix-tree usage by XArray in tcmu more than 1 year ago unfortunately got lost.
I rebased that work on latest tcmu and tested it.
--- drivers/target/target_core_user.c | 34 ++++++++++++++++------------------ 1 file changed, 16 insertions(+), 18 deletions(-)
drivers/target/target_core_user.c +16 -18
@@ -14,7 +14,6 @@#include <linux/vmalloc.h> #include <linux/uio_driver.h> #include <linux/xarray.h> -#include <linux/radix-tree.h> #include <linux/stringify.h> #include <linux/bitops.h> #include <linux/highmem.h>
@@ -143,7 +142,7 @@ struct tcmu_dev {uint32_t dbi_max;uint32_t dbi_thresh;unsigned long *data_bitmap; -struct radix_tree_root data_blocks; +struct xarray data_blocks;struct xarray commands;
@@ -500,13 +499,13 @@ static inline int tcmu_get_empty_block(struct tcmu_dev *udev, int prev_dbi, int *iov_cnt) { struct page *page; -int ret, dbi; +int dbi; dbi =find_first_zero_bit (udev->data_bitmap, udev->dbi_thresh);if (dbi == udev->dbi_thresh)return -1 ; - page =radix_tree_lookup (&udev->data_blocks, dbi); + page =xa_load (&udev->data_blocks, dbi);if (!page) {if (atomic_add_return (1 , &global_db_count) > tcmu_global_max_blocks)
@@ -517,8 +516,7 @@ static inline int tcmu_get_empty_block(struct tcmu_dev *udev,if (!page) goto err_alloc; - ret =radix_tree_insert (&udev->data_blocks, dbi, page); -if (ret) +if (xa_store (&udev->data_blocks, dbi, page,GFP_KERNEL )) goto err_insert; }
@@ -557,7 +555,7 @@ static int tcmu_get_empty_blocks(struct tcmu_dev *udev,static inline struct page *tcmu_get_block_page (struct tcmu_dev *udev,uint32_t dbi) { -return radix_tree_lookup (&udev->data_blocks, dbi); +return xa_load (&udev->data_blocks, dbi); }static inline void tcmu_free_cmd (struct tcmu_cmd *tcmu_cmd)
@@ -1561,7 +1559,7 @@ static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)timer_setup (&udev->qfull_timer, tcmu_qfull_timedout,0 );timer_setup (&udev->cmd_timer, tcmu_cmd_timedout,0 ); -INIT_RADIX_TREE (&udev->data_blocks,GFP_KERNEL ); +xa_init (&udev->data_blocks);return &udev->se_dev; }
@@ -1585,19 +1583,19 @@ static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)return -EINVAL ; } -static void tcmu_blocks_release (struct radix_tree_root *blocks, -int start,int end) +static void tcmu_blocks_release (struct xarray *blocks,unsigned long first, +unsigned long last) { -int i; +XA_STATE (xas, blocks, first);struct page *page; -for (i = start; i < end; i++) { - page =radix_tree_delete (blocks, i); -if (page) { -__free_page (page); -atomic_dec (&global_db_count); - } +xas_lock (&xas); +xas_for_each (&xas, page, last) { +xas_store (&xas,NULL ); +__free_page (page); +atomic_dec (&global_db_count); } +xas_unlock (&xas); }static void tcmu_remove_all_queued_tmr (struct tcmu_dev *udev)
@@ -2923,7 +2921,7 @@ static void find_free_blocks(void)unmap_mapping_range (udev->inode->i_mapping, off,0 ,1 );/* Release the block pages */ -tcmu_blocks_release (&udev->data_blocks, start, end); +tcmu_blocks_release (&udev->data_blocks, start, end -1 );mutex_unlock (&udev->cmdr_lock); total_freed += end - start;
Signature · 1 line
2.12.3
Quoted 10 lines · Bodo Stroesser
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro.
Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.
Quoted 7 lines
Bodo Stroesser (2):
scsi: target: tcmu: Replace IDR by XArray
scsi: target: tcmu: Replaceradix_treewith XArraydrivers/target/target_core_user.c | 64 +++++++++++++++++++-------------------- 1 file changed, 31 insertions(+), 33 deletions(-)
Looks ok to me.
I think in a separate patch we need to change:
+ if (xa_store(&udev->data_blocks, dbi, page, GFP_KERNEL))
goto err_insert;
to GFP_NOIO. There were some users doing tcm loop + tcmu and were hitting issues when GFP_KERNEL lead to write backs back on to the same tcmu device. We tried to change all the gfp flags but we missed that one, because it was hidden in the radix tree's xa_flags.
Bodo Stroesser
maintainerQuoted 28 lines · On 26.02.21 04:59, [email protected] wrote:
Quoted 10 lines · Bodo Stroesser
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro.
Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.Quoted 7 lines
Bodo Stroesser (2):
scsi: target: tcmu: Replace IDR by XArray
scsi: target: tcmu: Replaceradix_treewith XArraydrivers/target/target_core_user.c | 64 +++++++++++++++++++-------------------- 1 file changed, 31 insertions(+), 33 deletions(-)Looks ok to me.
Reviewed-by Mike Christie <[email protected]>I think in a separate patch we need to change:
(quoted diff) +1
-0+ if (xa_store(&udev->data_blocks, dbi, page, GFP_KERNEL)) goto err_insert;
Quoted 6 lines
to
GFP_NOIO. There were some users doing tcm loop + tcmu and were hitting issues whenGFP_KERNELlead to write backs back on to the same tcmu device. We tried to change all the gfp flags but we missed that one, because it was hidden in the radix tree'sxa_flags.
But then, for consistency reasons, shouldn't we change
+ if (xa_alloc(&udev->commands, &cmd_id, tcmu_cmd, XA_LIMIT(1, 0xffff),
+ GFP_NOWAIT) < 0) {
to GFP_NOIO also?
Additionally I have to change memory allocation in tcmu_tmr_notify from GFP_KERNEL to GFP_NOIO, since it happens while holding cmdr_lock mutex, which could also cause the problems you desribed.
Shouldn't we better change to new API memalloc_noio_save and memalloc_noio_restore in that new patch?
Quoted 36 lines · Bodo Stroesser
Quoted 25 lines · On 26.02.21 04:59, [email protected] wrote:
Quoted 10 lines · Bodo Stroesser
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro.
Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.Quoted 7 lines
Bodo Stroesser (2):
scsi: target: tcmu: Replace IDR by XArray
scsi: target: tcmu: Replaceradix_treewith XArray
drivers/target/target_core_user.c| 64 +++++++++++++++++++--------------------1 file changed, 31 insertions(+), 33 deletions(-)Looks ok to me.
Reviewed-by Mike Christie <[email protected]>I think in a separate patch we need to change:
Quoted 9 lines
+ if (
xa_store(&udev->data_blocks, dbi, page, GFP_KERNEL)) gotoerr_insert;to
GFP_NOIO. There were some users doing tcm loop + tcmu and were hitting issues whenGFP_KERNELlead to write backs back on to the same tcmu device. We tried to change all the gfp flags but we missed that one, because it was hidden in the radix tree'sxa_flags.
Quoted 6 lines
But then, for consistency reasons, shouldn't we change
+ if (
xa_alloc(&udev->commands, &cmd_id, tcmu_cmd, XA_LIMIT(1, 0xffff), + GFP_NOWAIT)< 0) {to
GFP_NOIOalso?
I think so, but am not sure. I've always wondered why we used GFP_NOWAIT and meant to test with different gfp values but didn't have time. It wouldn't hit the same issue I mentioned though.
Quoted 7 lines
Additionally I have to change memory allocation in
tcmu_tmr_notifyfromGFP_KERNELtoGFP_NOIO, since it happens while holdingcmdr_lockmutex, which could also cause the problems you desribed.Shouldn't we better change to new API
memalloc_noio_saveandmemalloc_noio_restorein that new patch?
I think it's your preference. I like to use the correct gfp when I can, so you can just look at that chunk of code and know it's right. If you put memalloc_noio_save in tcmu_queue_cmd then a couple functions down you have a GFP_KERNEL it's confusing when you are just searching the code.
I think memalloc_noio_save is handy in other places like the iscsi/nbd code since when calling into the network stack you can't control all the allocations sometimes.
Bodo Stroesser
maintainerQuoted 39 lines · Mike Christie
Quoted 38 lines · Bodo Stroesser
Quoted 25 lines · On 26.02.21 04:59, [email protected] wrote:
Quoted 10 lines · Bodo Stroesser
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro.
Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.Quoted 7 lines
Bodo Stroesser (2):
scsi: target: tcmu: Replace IDR by XArray
scsi: target: tcmu: Replaceradix_treewith XArraydrivers/target/target_core_user.c | 64 +++++++++++++++++++-------------------- 1 file changed, 31 insertions(+), 33 deletions(-)Looks ok to me.
Reviewed-by Mike Christie <[email protected]>I think in a separate patch we need to change:
Quoted 9 lines
+ if (
xa_store(&udev->data_blocks, dbi, page, GFP_KERNEL))
gotoerr_insert;to
GFP_NOIO. There were some users doing tcm loop + tcmu and were hitting issues whenGFP_KERNELlead to write backs back on to the same tcmu device. We tried to change all the gfp flags but we missed that one, because it was hidden in the radix tree'sxa_flags.But then, for consistency reasons, shouldn't we change
Quoted 9 lines
Quoted 4 lines
+ if (
xa_alloc(&udev->commands, &cmd_id, tcmu_cmd, XA_LIMIT(1, 0xffff), + GFP_NOWAIT)< 0) {to
GFP_NOIOalso?I think so, but am not sure. I've always wondered why we used
GFP_NOWAITand meant to test with different gfp values but didn't have time. It wouldn't hit the same issue I mentioned though.
I'll send a patch changing it to GFP_NOIO.
There is no reason to disallow sleeping. Just IO should be avoided as you said.
Quoted 5 lines
Quoted 5 lines
Additionally I have to change memory allocation in
tcmu_tmr_notifyfromGFP_KERNELtoGFP_NOIO, since it happens while holdingcmdr_lockmutex, which could also cause the problems you desribed.
Quoted 9 lines
Quoted 2 lines
Shouldn't we better change to new API
memalloc_noio_saveandmemalloc_noio_restorein that new patch?I think it's your preference. I like to use the correct gfp when I can, so you can just look at that chunk of code and know it's right. If you put
memalloc_noio_saveintcmu_queue_cmdthen a couple functions down you have aGFP_KERNELit's confusing when you are just searching the code.
Good point. I'll stay with GFP_NOIO.
Quoted 6 lines
I think
memalloc_noio_saveis handy in other places like the iscsi/nbd code since when calling into the network stack you can't control all the allocations sometimes.
Martin K. Petersen
AppliedBodo,
Quoted 6 lines
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro. Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.
Applied to 5.13/scsi-staging, thanks!
Signature · 1 line
Martin K. Petersen Oracle Linux Engineering
Martin K. Petersen
AppliedQuoted 11 lines · Bodo Stroesser
This small series is based on Martin's for-next.
Future patches will need something like the - meanwhile removed -
radix_tree_for_each_contigmacro.
Since general direction is to use xarray as replacement forradix_treeand IDR, instead of re-introducing or open coding the removed macro, with this series we switch over to xarray API. Based on xarray a future patch easily can implement an analog toradix_tree_for_each_contig.[...]
Applied to 5.13/scsi-queue, thanks!
[1/2] scsi: target: tcmu: Replace IDR by XArray
https://git.kernel.org/mkp/scsi/c/d3cbb743c362 [2/2] scsi: target: tcmu: Replace radix_tree with XArray https://git.kernel.org/mkp/scsi/c/f7c89771d07d
Signature · 1 line
Martin K. Petersen Oracle Linux Engineering