Re: [PATCH 3/6] nfs: Eliminate nfs_get_user_pages()

From: Chuck Lever
Date: Fri May 19 2006 - 15:19:13 EST


Andrew Morton wrote:
Chuck Lever <cel@xxxxxxxxxx> wrote:
Neil Brown observed that the kmalloc() in nfs_get_user_pages() is more
likely to fail if the I/O is large enough to require the allocation of more
than a single page to keep track of all the pinned pages in the user's
buffer.

Instead of tracking one large page array per dreq/iocb, track pages per
nfs_read/write_data, just like the cached I/O path does. An array for
pages is already allocated for us by nfs_readdata_alloc() (and the write
and commit equivalents).

This is also required for adding support for vectored I/O to the NFS direct
I/O path.

The original reason to pin the user buffer and allocate all the NFS data
structures before trying to schedule I/O was to ensure all needed resources
are allocated on the client before starting to send requests. This reduces
the chance that resource exhaustion on the client will cause a short read
or write.

On the other hand, for an application making very large application I/O
requests, this means that it will be nearly impossible for the application
to make forward progress on a resource-limited client.

Thus, moving the buffer pinning functionality into the I/O scheduling
loops should be good for scalability. The next patch will do the same for
NFS data structure allocation.

+static void nfs_release_user_pages(struct page **pages, int npages)
{
- int result = -ENOMEM;
- unsigned long page_count;
- size_t array_size;
-
- page_count = (user_addr + size + PAGE_SIZE - 1) >> PAGE_SHIFT;
- page_count -= user_addr >> PAGE_SHIFT;
-
- array_size = (page_count * sizeof(struct page *));
- *pages = kmalloc(array_size, GFP_KERNEL);
- if (*pages) {
- down_read(&current->mm->mmap_sem);
- result = get_user_pages(current, current->mm, user_addr,
- page_count, (rw == READ), 0,
- *pages, NULL);
- up_read(&current->mm->mmap_sem);
- if (result != page_count) {
- /*
- * If we got fewer pages than expected from
- * get_user_pages(), the user buffer runs off the
- * end of a mapping; return EFAULT.
- */
- if (result >= 0) {
- nfs_free_user_pages(*pages, result, 0);
- result = -EFAULT;
- } else
- kfree(*pages);
- *pages = NULL;
- }
- }
- return result;
+ int i;
+ for (i = 0; i < npages; i++)
+ page_cache_release(pages[i]);
}

If `npages' is negative, this does the right thing.

+ result = get_user_pages(current, current->mm, user_addr,
+ data->npages, 1, 0, data->pagevec, NULL);
+ up_read(&current->mm->mmap_sem);
+ if (unlikely(result < data->npages))
+ goto out_err;
...
+out_err:
+ nfs_release_user_pages(data->pagevec, result);

And `npages' can indeed be negative.

I fixed this by making all of these an "unsigned long". get_user_pages() returns an unsigned long result, so all these comparisons should always work correctly.

nfs_count_pages() now also returns an unsigned long, but I don't see how it is possible for it to compute a negative value.

So. No bug there, but the code is a little unobvious and fragile - if
someone were to alter a type then subtle bugs would happen.

Perhaps

if (result > 0)
nfs_release_user_pages(...);

would be cleaner. Or at least a loud comment in nfs_release_user_pages().

--
corporate: cel at netapp dot com
personal: chucklever at bigfoot dot com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/