2017-02-22 12:11 GMT-08:00 Jeff Layton <jlayton@xxxxxxxxxx>:
> On Fri, 2017-02-10 at 15:17 -0800, Pavel Shilovsky wrote:
>> Currently the code doesn't recognize asynchronous calls passed
>> by io_submit() and processes all calls synchronously. This is not
>> what kernel AIO expects. This patch improves this for reading
>> by introducing new async context that keeps track of all issued i/o
>> requests and moving a response collecting procedure to a separate
>> thread. This allows to return to the caller immediately for
>> asynchronous calls and call the iocb->ki_complete() once all requests
>> are completed. For synchronous calls the current thread simply
>> waits until all requests are completed.
>>
>> This improves reading performance of single threaded applications
>> with increasing of i/o queue depth size.
>>
>> Signed-off-by: Pavel Shilovsky <pshilov@xxxxxxxxxxxxx>
>> ---
>> fs/cifs/cifsglob.h | 18 ++++
>> fs/cifs/file.c | 273 +++++++++++++++++++++++++++++++++++++++++++++--------
>> 2 files changed, 252 insertions(+), 39 deletions(-)
>>
>> diff --git a/fs/cifs/cifsglob.h b/fs/cifs/cifsglob.h
>> index f9a9a12..80771ca 100644
>> --- a/fs/cifs/cifsglob.h
>> +++ b/fs/cifs/cifsglob.h
>> @@ -1109,6 +1109,23 @@ struct cifs_io_parms {
>> struct cifs_tcon *tcon;
>> };
>>
>
>
> My first thought here is that this patch is huge, and I'm having a
> difficult time following the changes that you're introducing. Have you
> considered breaking this up into smaller (and bisectable)changes. If
> there is a regression in here, tracking it down will be very difficuly
> with this patch as a single monolithic blob.
>
> My second thought here is that this is an awful lot of new
> infrastructure -- some more verbose explanation is warranted.
>
> Can we not just take a reference to the iocb and keep a pointer to it
> in cifs_readdata. Then when we go to fill pages or deal with an error,
> do the ki_complete? I guess each aio request can end up with multiple
> write SMBs on the wire, so maybe you do need something to aggregate all
> of that? I'm not sure I follow the rationale 100%...
Thank you for taking a look at this.
The are two parts here: adding the new async infrastructure and using
of this infrastructure for read code path. I can split the patch into
two but the first one won't make any difference to the code because it
would be just static functions that are not called anywhere.
The problem with async code path is that we don't have a direct access
to the memory under provided iovec argument from the kernel threads.
That's why we need to get underlying pages and map them once they are
needed.
You are right, one aio request can lead to multiple SMBs and we need
to aggregate the results. That's why cifs_aio_ctx structure was
introduced. Every readdata keeps a reference to this context and
populate the pages with data in the correct order.
>
>> +struct cifs_aio_ctx {
>> + struct kref refcount;
>> + struct list_head list;
>> + struct mutex aio_mutex;
>> + struct completion done;
>> + struct iov_iter iter;
>> + struct kiocb *iocb;
>> + struct cifsFileInfo *cfile;
>> + struct page **pages;
>> + struct bio_vec *bv;
>> + unsigned int npages;
>> + ssize_t rc;
>> + unsigned int len;
>> + unsigned int total_len;
>> + bool should_dirty;
>> +};
>> +
>
>> struct cifs_readdata;
>>
>> /* asynchronous read support */
>> @@ -1118,6 +1135,7 @@ struct cifs_readdata {
>> struct completion done;
>> struct cifsFileInfo *cfile;
>> struct address_space *mapping;
>> + struct cifs_aio_ctx *ctx;
>> __u64 offset;
>> unsigned int bytes;
>> unsigned int got_bytes;
>> diff --git a/fs/cifs/file.c b/fs/cifs/file.c
>> index 98dc842..6ceeed2 100644
>> --- a/fs/cifs/file.c
>> +++ b/fs/cifs/file.c
>> @@ -33,6 +33,7 @@
>> #include <linux/mount.h>
>> #include <linux/slab.h>
>> #include <linux/swap.h>
>> +#include <linux/vmalloc.h>
>> #include <asm/div64.h>
>> #include "cifsfs.h"
>> #include "cifspdu.h"
>> @@ -2410,6 +2411,108 @@ int cifs_flush(struct file *file, fl_owner_t id)
>> return rc;
>> }
>>
>> +static inline struct cifs_aio_ctx *
>> +cifs_aio_ctx_alloc(void)
>> +{
>> + struct cifs_aio_ctx *ctx;
>> +
>> + ctx = kzalloc(sizeof(struct cifs_aio_ctx), GFP_KERNEL);
>> + if (!ctx)
>> + return NULL;
>> +
>> + INIT_LIST_HEAD(&ctx->list);
>> + mutex_init(&ctx->aio_mutex);
>> + init_completion(&ctx->done);
>> + kref_init(&ctx->refcount);
>> + return ctx;
>> +}
>> +
>> +static inline void
>> +cifs_aio_ctx_release(struct kref *refcount)
>> +{
>> + struct cifs_aio_ctx *ctx = container_of(refcount,
>> + struct cifs_aio_ctx, refcount);
>> +
>> + cifsFileInfo_put(ctx->cfile);
>> + kvfree(ctx->bv);
>> + kfree(ctx);
>> +}
>> +
>> +static int
>> +setup_aio_ctx_iter(struct cifs_aio_ctx *ctx, struct iov_iter *iter, int rw)
>> +{
>> + ssize_t rc;
>> + unsigned int cur_npages;
>> + unsigned int npages = 0;
>> + unsigned int i;
>> + size_t len;
>> + size_t count = iov_iter_count(iter);
>> + unsigned int saved_len;
>> + size_t start;
>> + unsigned int max_pages = iov_iter_npages(iter, INT_MAX);
>> + struct page **pages;
>> + struct bio_vec *bv;
>> +
>> + if (iter->type & ITER_KVEC) {
>> + memcpy(&ctx->iter, iter, sizeof(struct iov_iter));
>> + ctx->len = count;
>> + iov_iter_advance(iter, count);
>> + return 0;
>> + }
>> +
>> + bv = kmalloc_array(max_pages, sizeof(struct bio_vec), GFP_KERNEL);
>> + if (!bv) {
>> + bv = vmalloc(max_pages * sizeof(struct bio_vec));
>> + if (!bv)
>> + return -ENOMEM;
>> + }
>> +
>> + saved_len = count;
>> +
>> + while (count && npages < max_pages) {
>> + rc = iov_iter_get_pages_alloc(iter, &pages, count, &start);
>> + if (rc < 0) {
>> + cifs_dbg(VFS, "couldn't get user pages (rc=%zd)\n", rc);
>> + break;
>> + }
>> +
>> + if (rc > count) {
>> + cifs_dbg(VFS, "get pages rc=%zd more than %zu\n", rc,
>> + count);
>> + break;
>> + }
>> +
>> + iov_iter_advance(iter, rc);
>> + count -= rc;
>> + rc += start;
>> + cur_npages = (rc + PAGE_SIZE - 1) / PAGE_SIZE;
>> +
>> + if (npages + cur_npages > max_pages) {
>> + cifs_dbg(VFS, "out of vec array capacity (%u vs %u)\n",
>> + npages + cur_npages, max_pages);
>> + break;
>> + }
>> +
>> + for (i = 0; i < cur_npages; i++) {
>> + len = rc > PAGE_SIZE ? PAGE_SIZE : rc;
>> + bv[npages + i].bv_page = pages[i];
>> + bv[npages + i].bv_offset = start;
>> + bv[npages + i].bv_len = len - start;
>> + rc -= len;
>> + start = 0;
>> + }
>> +
>> + npages += cur_npages;
>> + kvfree(pages);
>> + }
>> +
>> + ctx->bv = bv;
>> + ctx->len = saved_len - count;
>> + ctx->npages = npages;
>> + iov_iter_bvec(&ctx->iter, ITER_BVEC | rw, ctx->bv, npages, ctx->len);
>> + return 0;
>> +}
>> +
>> static int
>> cifs_write_allocate_pages(struct page **pages, unsigned long num_pages)
>> {
>> @@ -2859,6 +2962,7 @@ cifs_uncached_readdata_release(struct kref *refcount)
>> struct cifs_readdata, refcount);
>> unsigned int i;
>>
>> + kref_put(&rdata->ctx->refcount, cifs_aio_ctx_release);
>> for (i = 0; i < rdata->nr_pages; i++) {
>> put_page(rdata->pages[i]);
>> rdata->pages[i] = NULL;
>> @@ -2900,6 +3004,8 @@ cifs_readdata_to_iov(struct cifs_readdata *rdata, struct iov_iter *iter)
>> return remaining ? -EFAULT : 0;
>> }
>>
>> +static void collect_uncached_read_data(struct cifs_aio_ctx *ctx);
>> +
>> static void
>> cifs_uncached_readv_complete(struct work_struct *work)
>> {
>> @@ -2907,6 +3013,8 @@ cifs_uncached_readv_complete(struct work_struct *work)
>> struct cifs_readdata, work);
>>
>> complete(&rdata->done);
>> + collect_uncached_read_data(rdata->ctx);
>> + /* the below call can possibly free the last ref to aio ctx */
>> kref_put(&rdata->refcount, cifs_uncached_readdata_release);
>> }
>>
>> @@ -2973,7 +3081,8 @@ cifs_uncached_copy_into_pages(struct TCP_Server_Info *server,
>>
>> static int
>> cifs_send_async_read(loff_t offset, size_t len, struct cifsFileInfo *open_file,
>> - struct cifs_sb_info *cifs_sb, struct list_head *rdata_list)
>> + struct cifs_sb_info *cifs_sb, struct list_head *rdata_list,
>> + struct cifs_aio_ctx *ctx)
>> {
>> struct cifs_readdata *rdata;
>> unsigned int npages, rsize, credits;
>> @@ -3020,6 +3129,8 @@ cifs_send_async_read(loff_t offset, size_t len, struct cifsFileInfo *open_file,
>> rdata->read_into_pages = cifs_uncached_read_into_pages;
>> rdata->copy_into_pages = cifs_uncached_copy_into_pages;
>> rdata->credits = credits;
>> + rdata->ctx = ctx;
>> + kref_get(&ctx->refcount);
>>
>> if (!rdata->cfile->invalidHandle ||
>> !cifs_reopen_file(rdata->cfile, true))
>> @@ -3042,50 +3153,37 @@ cifs_send_async_read(loff_t offset, size_t len, struct cifsFileInfo *open_file,
>> return rc;
>> }
>>
>> -ssize_t cifs_user_readv(struct kiocb *iocb, struct iov_iter *to)
>> +static void
>> +collect_uncached_read_data(struct cifs_aio_ctx *ctx)
>> {
>> - struct file *file = iocb->ki_filp;
>> - ssize_t rc;
>> - size_t len;
>> - ssize_t total_read = 0;
>> - loff_t offset = iocb->ki_pos;
>> + struct cifs_readdata *rdata, *tmp;
>> + struct iov_iter *to = &ctx->iter;
>> struct cifs_sb_info *cifs_sb;
>> struct cifs_tcon *tcon;
>> - struct cifsFileInfo *open_file;
>> - struct cifs_readdata *rdata, *tmp;
>> - struct list_head rdata_list;
>> -
>> - len = iov_iter_count(to);
>> - if (!len)
>> - return 0;
>> -
>> - INIT_LIST_HEAD(&rdata_list);
>> - cifs_sb = CIFS_FILE_SB(file);
>> - open_file = file->private_data;
>> - tcon = tlink_tcon(open_file->tlink);
>> -
>> - if (!tcon->ses->server->ops->async_readv)
>> - return -ENOSYS;
>> + unsigned int i;
>> + int rc;
>>
>> - if ((file->f_flags & O_ACCMODE) == O_WRONLY)
>> - cifs_dbg(FYI, "attempting read on write only file instance\n");
>> + tcon = tlink_tcon(ctx->cfile->tlink);
>> + cifs_sb = CIFS_SB(ctx->cfile->dentry->d_sb);
>>
>> - rc = cifs_send_async_read(offset, len, open_file, cifs_sb, &rdata_list);
>> + mutex_lock(&ctx->aio_mutex);
>>
>> - /* if at least one read request send succeeded, then reset rc */
>> - if (!list_empty(&rdata_list))
>> - rc = 0;
>> + if (list_empty(&ctx->list)) {
>> + mutex_unlock(&ctx->aio_mutex);
>> + return;
>> + }
>>
>> - len = iov_iter_count(to);
>> + rc = ctx->rc;
>> /* the loop below should proceed in the order of increasing offsets */
>> again:
>> - list_for_each_entry_safe(rdata, tmp, &rdata_list, list) {
>> + list_for_each_entry_safe(rdata, tmp, &ctx->list, list) {
>> if (!rc) {
>> - /* FIXME: freezable sleep too? */
>> - rc = wait_for_completion_killable(&rdata->done);
>> - if (rc)
>> - rc = -EINTR;
>> - else if (rdata->result == -EAGAIN) {
>> + if (!try_wait_for_completion(&rdata->done)) {
>> + mutex_unlock(&ctx->aio_mutex);
>> + return;
>> + }
>> +
>
> The coarse-grained aio_mutex here is particularly nasty, and will be a
> pain to rip out later if it turns out to be a performance bottleneck.
> It's held over almost all of this new code, so it's not clear what it
> protects. Is it really necessary, or could you get away with some
> lighter-weight synchronization -- e.g. spinlocks?
The mutex is used to prevent parallel access to a context internal
fields (ctx->list, ctx->rc) from multiple readdata completion
routines. Since we need to populate the pages in the correct order it
shouldn't affect the performance. Also we can retry the request that
may split it into several ones (which need to be placed into correct
positions in the list) and makes things more complicated if we want to
use spinlocks. Note that we don't want to retry the request if the
previous ones (in the order) failed. If case of no error, we won't
hold the mutex for a long time.
>
>
>> + if (rdata->result == -EAGAIN) {
>> /* resend call if it's a retryable error */
>> struct list_head tmp_list;
>> unsigned int got_bytes = rdata->got_bytes;
>> @@ -3111,9 +3209,9 @@ ssize_t cifs_user_readv(struct kiocb *iocb, struct iov_iter *to)
>> rdata->offset + got_bytes,
>> rdata->bytes - got_bytes,
>> rdata->cfile, cifs_sb,
>> - &tmp_list);
>> + &tmp_list, ctx);
>>
>> - list_splice(&tmp_list, &rdata_list);
>> + list_splice(&tmp_list, &ctx->list);
>>
>> kref_put(&rdata->refcount,
>> cifs_uncached_readdata_release);
>> @@ -3131,14 +3229,111 @@ ssize_t cifs_user_readv(struct kiocb *iocb, struct iov_iter *to)
>> kref_put(&rdata->refcount, cifs_uncached_readdata_release);
>> }
>>
>> - total_read = len - iov_iter_count(to);
>> + for (i = 0; i < ctx->npages; i++) {
>> + if (ctx->should_dirty)
>> + set_page_dirty(ctx->bv[i].bv_page);
>> + put_page(ctx->bv[i].bv_page);
>> + }
>> +
>> + ctx->total_len = ctx->len - iov_iter_count(to);
>>
>> - cifs_stats_bytes_read(tcon, total_read);
>> + cifs_stats_bytes_read(tcon, ctx->total_len);
>>
>> /* mask nodata case */
>> if (rc == -ENODATA)
>> rc = 0;
>>
>> + ctx->rc = (rc == 0) ? ctx->total_len : rc;
>> +
>> + mutex_unlock(&ctx->aio_mutex);
>> +
>> + if (ctx->iocb && ctx->iocb->ki_complete)
>> + ctx->iocb->ki_complete(ctx->iocb, ctx->rc, 0);
>> + else
>> + complete(&ctx->done);
>> +}
>> +
>
> I haven't gone to go dig around in the AIO code, but what guarantees
> that ctx->iocb still points to valid memory once you get here. This
> will be called in workqueue context when the read reply comes in?
ctx->iocb points to iocb argument which is allocated on the heap for
async code path. For sync code path iocb is stack allocated, so
ctx->iocb is NULL and we call complete().
>
> Is it possible that this kiocb could be freed (via aio_cancel perhaps?)
> while there are still requests in flight?
We do not set ki_cancel(), so an async job can't be cancelled. I
didn't find any other cases when it can be freed. Also I looked at
other filesystem modules (e.g. FUSE), it seems like they reply on this
too.
>> +ssize_t cifs_user_readv(struct kiocb *iocb, struct iov_iter *to)
>> +{
>> + struct file *file = iocb->ki_filp;
>> + ssize_t rc;
>> + size_t len;
>> + ssize_t total_read = 0;
>> + loff_t offset = iocb->ki_pos;
>> + struct cifs_sb_info *cifs_sb;
>> + struct cifs_tcon *tcon;
>> + struct cifsFileInfo *cfile;
>> + struct cifs_aio_ctx *ctx;
>> +
>> + len = iov_iter_count(to);
>> + if (!len)
>> + return 0;
>> +
>> + cifs_sb = CIFS_FILE_SB(file);
>> + cfile = file->private_data;
>> + tcon = tlink_tcon(cfile->tlink);
>> +
>> + if (!tcon->ses->server->ops->async_readv)
>> + return -ENOSYS;
>> +
>> + if ((file->f_flags & O_ACCMODE) == O_WRONLY)
>> + cifs_dbg(FYI, "attempting read on write only file instance\n");
>> +
>> + ctx = cifs_aio_ctx_alloc();
>> + if (!ctx)
>> + return -ENOMEM;
>> +
>> + ctx->cfile = cifsFileInfo_get(cfile);
>> +
>> + if (!is_sync_kiocb(iocb))
>> + ctx->iocb = iocb;
>> +
>> + if (to->type & ITER_IOVEC)
>> + ctx->should_dirty = true;
>> +
>> + rc = setup_aio_ctx_iter(ctx, to, READ);
>> + if (rc) {
>> + kref_put(&ctx->refcount, cifs_aio_ctx_release);
>> + return rc;
>> + }
>> +
>> + len = ctx->len;
>> +
>> + /* grab a lock here due to read response handlers can access ctx */
>> + mutex_lock(&ctx->aio_mutex);
We need to hold the mutex here because we populate ctx->list here and
during the sending procedure the responses can be received and
processed.
>> +
>> + rc = cifs_send_async_read(offset, len, cfile, cifs_sb, &ctx->list, ctx);
>> +
>> + /* if at least one read request send succeeded, then reset rc */
>> + if (!list_empty(&ctx->list))
>> + rc = 0;
>> +
>> + mutex_unlock(&ctx->aio_mutex);
>> +
>> + if (rc) {
>> + kref_put(&ctx->refcount, cifs_aio_ctx_release);
>> + return rc;
>> + }
>> +
>> + if (!is_sync_kiocb(iocb)) {
>> + kref_put(&ctx->refcount, cifs_aio_ctx_release);
>> + return -EIOCBQUEUED;
>> + }
>> +
>> + /* FIXME: freezable sleep too? */
>> + rc = wait_for_completion_killable(&ctx->done);
>> + if (rc) {
>> + mutex_lock(&ctx->aio_mutex);
>> + ctx->rc = rc = -EINTR;
>> + total_read = ctx->total_len;
>> + mutex_unlock(&ctx->aio_mutex);
>> + } else {
>> + rc = ctx->rc;
>> + total_read = ctx->total_len;
>> + }
>> +
>> + kref_put(&ctx->refcount, cifs_aio_ctx_release);
>> +
>> if (total_read) {
>> iocb->ki_pos += total_read;
>> return total_read;
>
> --
> Jeff Layton <jlayton@xxxxxxxxxx>
--
Best regards,
Pavel Shilovsky
--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
