2023-04-17 06:43:10

by Yue Hu

[permalink] [raw]
Subject: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend

From: Yue Hu <[email protected]>

The icur field is only used in z_erofs_try_inplace_io(). Let's just use
a local variable instead. And no need to check if the pcluster is inline
when setting icur since inline page cannot be used for inplace I/O.

Signed-off-by: Yue Hu <[email protected]>
---
fs/erofs/zdata.c | 13 +++++--------
1 file changed, 5 insertions(+), 8 deletions(-)

diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
index f759152feffa..f8bf2b227942 100644
--- a/fs/erofs/zdata.c
+++ b/fs/erofs/zdata.c
@@ -554,9 +554,6 @@ struct z_erofs_decompress_frontend {
/* used for applying cache strategy on the fly */
bool backmost;
erofs_off_t headoffset;
-
- /* a pointer used to pick up inplace I/O pages */
- unsigned int icur;
};

#define DECOMPRESS_FRONTEND_INIT(__i) { \
@@ -707,11 +704,13 @@ static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
struct z_erofs_bvec *bvec)
{
struct z_erofs_pcluster *const pcl = fe->pcl;
+ /* file-backed online pages are traversed in reverse order */
+ unsigned int icur = pcl->pclusterpages;

- while (fe->icur > 0) {
- if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
+ while (icur > 0) {
+ if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
NULL, bvec->page)) {
- pcl->compressed_bvecs[fe->icur] = *bvec;
+ pcl->compressed_bvecs[icur] = *bvec;
return true;
}
}
@@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
}
z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
- /* since file-backed online pages are traversed in reverse order */
- fe->icur = z_erofs_pclusterpages(fe->pcl);
return 0;
}

--
2.17.1


2023-04-17 06:56:08

by Gao Xiang

[permalink] [raw]
Subject: Re: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend



On 2023/4/17 14:41, Yue Hu wrote:
> From: Yue Hu <[email protected]>
>
> The icur field is only used in z_erofs_try_inplace_io(). Let's just use
> a local variable instead. And no need to check if the pcluster is inline
> when setting icur since inline page cannot be used for inplace I/O.
>
> Signed-off-by: Yue Hu <[email protected]>

Nope, it's a behavior change.
Other users could feed more inplace I/O pages before I/O submission
to reduce memory consumption, it's common when applying stress model
in low memory scenarios.

Thanks,
Gao Xiang

> ---
> fs/erofs/zdata.c | 13 +++++--------
> 1 file changed, 5 insertions(+), 8 deletions(-)
>
> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
> index f759152feffa..f8bf2b227942 100644
> --- a/fs/erofs/zdata.c
> +++ b/fs/erofs/zdata.c
> @@ -554,9 +554,6 @@ struct z_erofs_decompress_frontend {
> /* used for applying cache strategy on the fly */
> bool backmost;
> erofs_off_t headoffset;
> -
> - /* a pointer used to pick up inplace I/O pages */
> - unsigned int icur;
> };
>
> #define DECOMPRESS_FRONTEND_INIT(__i) { \
> @@ -707,11 +704,13 @@ static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
> struct z_erofs_bvec *bvec)
> {
> struct z_erofs_pcluster *const pcl = fe->pcl;
> + /* file-backed online pages are traversed in reverse order */
> + unsigned int icur = pcl->pclusterpages;
>
> - while (fe->icur > 0) {
> - if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
> + while (icur > 0) {
> + if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
> NULL, bvec->page)) {
> - pcl->compressed_bvecs[fe->icur] = *bvec;
> + pcl->compressed_bvecs[icur] = *bvec;
> return true;
> }
> }
> @@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
> }
> z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
> Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
> - /* since file-backed online pages are traversed in reverse order */
> - fe->icur = z_erofs_pclusterpages(fe->pcl);
> return 0;
> }
>

2023-04-17 07:08:55

by Gao Xiang

[permalink] [raw]
Subject: Re: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend



On 2023/4/17 14:52, Gao Xiang wrote:
>
>
> On 2023/4/17 14:41, Yue Hu wrote:
>> From: Yue Hu <[email protected]>
>>
>> The icur field is only used in z_erofs_try_inplace_io(). Let's just use
>> a local variable instead. And no need to check if the pcluster is inline
>> when setting icur since inline page cannot be used for inplace I/O.

Where do we check if the pcluster is inline?

>>
>> Signed-off-by: Yue Hu <[email protected]>
>
> Nope, it's a behavior change.
> Other users could feed more inplace I/O pages before I/O submission
> to reduce memory consumption, it's common when applying stress model
> in low memory scenarios.

Oh, I misread it. I think it can be done in this way although
each following users will now rescan the whole array all the time.

Reviewed-by: Gao Xiang <[email protected]>

>
> Thanks,
> Gao Xiang
>
>> ---
>>   fs/erofs/zdata.c | 13 +++++--------
>>   1 file changed, 5 insertions(+), 8 deletions(-)
>>
>> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
>> index f759152feffa..f8bf2b227942 100644
>> --- a/fs/erofs/zdata.c
>> +++ b/fs/erofs/zdata.c
>> @@ -554,9 +554,6 @@ struct z_erofs_decompress_frontend {
>>       /* used for applying cache strategy on the fly */
>>       bool backmost;
>>       erofs_off_t headoffset;
>> -
>> -    /* a pointer used to pick up inplace I/O pages */
>> -    unsigned int icur;
>>   };
>>   #define DECOMPRESS_FRONTEND_INIT(__i) { \
>> @@ -707,11 +704,13 @@ static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
>>                      struct z_erofs_bvec *bvec)
>>   {
>>       struct z_erofs_pcluster *const pcl = fe->pcl;
>> +    /* file-backed online pages are traversed in reverse order */

Although please help refine the comment below:

/* scan & fill inplace I/O pages in the reverse order */

Thanks,
Gao Xiang

>> +    unsigned int icur = pcl->pclusterpages;
>> -    while (fe->icur > 0) {
>> -        if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
>> +    while (icur > 0) {
>> +        if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
>>                    NULL, bvec->page)) {
>> -            pcl->compressed_bvecs[fe->icur] = *bvec;
>> +            pcl->compressed_bvecs[icur] = *bvec;
>>               return true;
>>           }
>>       }
>> @@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
>>       }
>>       z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
>>                   Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
>> -    /* since file-backed online pages are traversed in reverse order */
>> -    fe->icur = z_erofs_pclusterpages(fe->pcl);
>>       return 0;
>>   }

2023-04-17 07:10:16

by Yue Hu

[permalink] [raw]
Subject: Re: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend

On Mon, 17 Apr 2023 15:00:25 +0800
Gao Xiang <[email protected]> wrote:

> On 2023/4/17 14:52, Gao Xiang wrote:
> >
> >
> > On 2023/4/17 14:41, Yue Hu wrote:
> >> From: Yue Hu <[email protected]>
> >>
> >> The icur field is only used in z_erofs_try_inplace_io(). Let's just use
> >> a local variable instead. And no need to check if the pcluster is inline
> >> when setting icur since inline page cannot be used for inplace I/O.
>
> Where do we check if the pcluster is inline?

/* since file-backed online pages are traversed in reverse order */
fe->icur = z_erofs_pclusterpages(fe->pcl);


static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)
{
if (z_erofs_is_inline_pcluster(pcl))
return 1;
[...]

>
> >>
> >> Signed-off-by: Yue Hu <[email protected]>
> >
> > Nope, it's a behavior change.
> > Other users could feed more inplace I/O pages before I/O submission
> > to reduce memory consumption, it's common when applying stress model
> > in low memory scenarios.
>
> Oh, I misread it. I think it can be done in this way although
> each following users will now rescan the whole array all the time.
>
> Reviewed-by: Gao Xiang <[email protected]>
>
> >
> > Thanks,
> > Gao Xiang
> >
> >> ---
> >> ?0?2 fs/erofs/zdata.c | 13 +++++--------
> >> ?0?2 1 file changed, 5 insertions(+), 8 deletions(-)
> >>
> >> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c
> >> index f759152feffa..f8bf2b227942 100644
> >> --- a/fs/erofs/zdata.c
> >> +++ b/fs/erofs/zdata.c
> >> @@ -554,9 +554,6 @@ struct z_erofs_decompress_frontend {
> >> ?0?2?0?2?0?2?0?2?0?2 /* used for applying cache strategy on the fly */
> >> ?0?2?0?2?0?2?0?2?0?2 bool backmost;
> >> ?0?2?0?2?0?2?0?2?0?2 erofs_off_t headoffset;
> >> -
> >> -?0?2?0?2?0?2 /* a pointer used to pick up inplace I/O pages */
> >> -?0?2?0?2?0?2 unsigned int icur;
> >> ?0?2 };
> >> ?0?2 #define DECOMPRESS_FRONTEND_INIT(__i) { \
> >> @@ -707,11 +704,13 @@ static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
> >> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 struct z_erofs_bvec *bvec)
> >> ?0?2 {
> >> ?0?2?0?2?0?2?0?2?0?2 struct z_erofs_pcluster *const pcl = fe->pcl;
> >> +?0?2?0?2?0?2 /* file-backed online pages are traversed in reverse order */
>
> Although please help refine the comment below:
>
> /* scan & fill inplace I/O pages in the reverse order */

Ok, will refine it in v2.

>
> Thanks,
> Gao Xiang
>
> >> +?0?2?0?2?0?2 unsigned int icur = pcl->pclusterpages;
> >> -?0?2?0?2?0?2 while (fe->icur > 0) {
> >> -?0?2?0?2?0?2?0?2?0?2?0?2?0?2 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
> >> +?0?2?0?2?0?2 while (icur > 0) {
> >> +?0?2?0?2?0?2?0?2?0?2?0?2?0?2 if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
> >> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 NULL, bvec->page)) {
> >> -?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 pcl->compressed_bvecs[fe->icur] = *bvec;
> >> +?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 pcl->compressed_bvecs[icur] = *bvec;
> >> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 return true;
> >> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 }
> >> ?0?2?0?2?0?2?0?2?0?2 }
> >> @@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
> >> ?0?2?0?2?0?2?0?2?0?2 }
> >> ?0?2?0?2?0?2?0?2?0?2 z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
> >> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
> >> -?0?2?0?2?0?2 /* since file-backed online pages are traversed in reverse order */
> >> -?0?2?0?2?0?2 fe->icur = z_erofs_pclusterpages(fe->pcl);
> >> ?0?2?0?2?0?2?0?2?0?2 return 0;
> >> ?0?2 }

2023-04-17 07:16:18

by Gao Xiang

[permalink] [raw]
Subject: Re: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend



On 2023/4/17 15:15, Yue Hu wrote:
> On Mon, 17 Apr 2023 15:00:25 +0800
> Gao Xiang <[email protected]> wrote:
>

..

>>
>> Although please help refine the comment below:
>>
>> /* scan & fill inplace I/O pages in the reverse order */
>
> Ok, will refine it in v2.

I rethink this, I don't want to go far in this way, and this makes a
O(n) scan into O(n^2) when a single inplace I/O page is added.

So sorry, I don't think it's a good way, although I also don't think
`icur` is a good name and we might need to find a better name.

Thanks,
Gao Xiang

>
>>
>> Thanks,
>> Gao Xiang
>>
>>>> +    unsigned int icur = pcl->pclusterpages;
>>>> -    while (fe->icur > 0) {
>>>> -        if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
>>>> +    while (icur > 0) {
>>>> +        if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
>>>>                    NULL, bvec->page)) {
>>>> -            pcl->compressed_bvecs[fe->icur] = *bvec;
>>>> +            pcl->compressed_bvecs[icur] = *bvec;
>>>>               return true;
>>>>           }
>>>>       }
>>>> @@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
>>>>       }
>>>>       z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
>>>>                   Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
>>>> -    /* since file-backed online pages are traversed in reverse order */
>>>> -    fe->icur = z_erofs_pclusterpages(fe->pcl);
>>>>       return 0;
>>>>   }

2023-04-17 07:38:41

by Yue Hu

[permalink] [raw]
Subject: Re: [PATCH] erofs: remove unneeded icur field from struct z_erofs_decompress_frontend

On Mon, 17 Apr 2023 15:12:31 +0800
Gao Xiang <[email protected]> wrote:

> On 2023/4/17 15:15, Yue Hu wrote:
> > On Mon, 17 Apr 2023 15:00:25 +0800
> > Gao Xiang <[email protected]> wrote:
> >
>
> ..
>
> >>
> >> Although please help refine the comment below:
> >>
> >> /* scan & fill inplace I/O pages in the reverse order */
> >
> > Ok, will refine it in v2.
>
> I rethink this, I don't want to go far in this way, and this makes a
> O(n) scan into O(n^2) when a single inplace I/O page is added.

Yeah, i misread that, it should be global before submission, just ignore the change.

>
> So sorry, I don't think it's a good way, although I also don't think
> `icur` is a good name and we might need to find a better name.
>
> Thanks,
> Gao Xiang
>
> >
> >>
> >> Thanks,
> >> Gao Xiang
> >>
> >>>> +?0?2?0?2?0?2 unsigned int icur = pcl->pclusterpages;
> >>>> -?0?2?0?2?0?2 while (fe->icur > 0) {
> >>>> -?0?2?0?2?0?2?0?2?0?2?0?2?0?2 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
> >>>> +?0?2?0?2?0?2 while (icur > 0) {
> >>>> +?0?2?0?2?0?2?0?2?0?2?0?2?0?2 if (!cmpxchg(&pcl->compressed_bvecs[--icur].page,
> >>>> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 NULL, bvec->page)) {
> >>>> -?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 pcl->compressed_bvecs[fe->icur] = *bvec;
> >>>> +?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 pcl->compressed_bvecs[icur] = *bvec;
> >>>> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 return true;
> >>>> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 }
> >>>> ?0?2?0?2?0?2?0?2?0?2 }
> >>>> @@ -877,8 +876,6 @@ static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
> >>>> ?0?2?0?2?0?2?0?2?0?2 }
> >>>> ?0?2?0?2?0?2?0?2?0?2 z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
> >>>> ?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2?0?2 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
> >>>> -?0?2?0?2?0?2 /* since file-backed online pages are traversed in reverse order */
> >>>> -?0?2?0?2?0?2 fe->icur = z_erofs_pclusterpages(fe->pcl);
> >>>> ?0?2?0?2?0?2?0?2?0?2 return 0;
> >>>> ?0?2 }