Hi Linus, (and list members)
The patch below addresses all known functional issues with the ramdisk
and passes all tests I can think of.
What it doesn't solve, is a (design?) weakness whereby the ramdisk can
cause swap looping if it's trying to grow_buffers() to dynamically
allocate blocks during high memory pressure times. This is exposed
by swapping madly (ultimate test) to a ramdisk. It passes the test
just fine iff I populate the ramdisk (dd) prior to mkswap/swapon on
the device. I'm running the result of make -j3 bzImage MAKE='make -j3'
with the ramswap being enabled/disabled/enabled repeatedly during build,
migrating it's content to a lower priority swap partiton and back.
Changes:
Unconditionally retrieve/create/populate buffers during rd_request.
This resolves a problem whereby a ramdisk filled with an fs image
then wiped and re-filled has files full of zeros. My understanding
of the buffer cache + page cache == unified cache is too shaky to
explain this at all. My _impression_ is that the total cache state
must be stored in the ramdisk (lives in buffer cache portion of total)
for it to work properly (ie aliases=state problem). This is as shaky
as explanations come, but that's what I'm trying to do, and it works.
(Tigran tested/confirmed that it resolves that problem)
Added igrab()/iput() to protect the ramdisk inodes from prune_icache()
when ramdisks are present but inactive.
Added setup option for blocksize when ramdisk is in kernel such that
options are the same as what's available as a module. This is needed
if you want to swap to ramdisk when it's in kernel.. a valid way to
test swap performance on a machine with terrible/old disk drives :)
Tiny cleanups.
If anything is horrible from the OS designer viewpoint, punt me in the
right direction and I'll try to do better. Patch is below.
Thanks for your time,
-Mike
--- linux-2.3.50.virgin/drivers/block/rd.c Fri Mar 3 06:28:01 2000
+++ linux-2.3.49.test/drivers/block/rd.c Wed Mar 8 16:36:24 2000
@@ -99,6 +99,7 @@
static int rd_blocksizes[NUM_RAMDISKS]; /* Size of 1024 byte blocks :) */
static int rd_kbsize[NUM_RAMDISKS]; /* Size in blocks of 1024 bytes */
static devfs_handle_t devfs_handle = NULL;
+static struct inode *rd_inode[NUM_RAMDISKS]; /* Protected device inodes */
/*
* Parameters for the boot-loading of the RAM disk. These are set by
@@ -169,11 +170,18 @@
return ramdisk_size(str);
}
+static int __init ramdisk_blocksize(char *str)
+{
+ rd_blocksize = simple_strtol(str,NULL,0);
+ return 1;
+}
+
__setup("ramdisk_start=", ramdisk_start_setup);
__setup("load_ramdisk=", load_ramdisk);
__setup("prompt_ramdisk=", prompt_ramdisk);
__setup("ramdisk=", ramdisk_size);
__setup("ramdisk_size=", ramdisk_size2);
+__setup("ramdisk_blocksize=", ramdisk_blocksize);
#endif
@@ -216,67 +224,16 @@
goto repeat;
}
- /*
- * This has become somewhat more complicated with the addition of
- * the page cache. The problem is that in some cases the furnished
- * buffer is "real", i.e., part of the existing ramdisk, while in
- * others it is "unreal", e.g., part of a page. In the first case
- * not much needs to be done, while in the second, some kind of
- * transfer is needed.
- *
- * The two cases are distinguished here by checking whether the
- * real buffer is already in the buffer cache, and whether it is
- * the same as the one supplied.
- *
- * There are three cases with read/write to consider:
- *
- * 1. Supplied buffer matched one in the buffer cache:
- * Read - Clear the buffer, as it wasn't already valid.
- * Write - Mark the buffer as "Protected".
- *
- * 2. Supplied buffer mismatched one in the buffer cache:
- * Read - Copy the data from the buffer cache entry.
- * Write - Copy the data to the buffer cache entry.
- *
- * 3 No buffer cache entry existed:
- * Read - Clear the supplied buffer, but do not create a real
- * one.
- * Write - Create a real buffer, copy the data to it, and mark
- * it as "Protected".
- *
- * NOTE: There seems to be some schizophrenia here - the logic
- * using "len" seems to assume arbitrary request lengths, while
- * the "protect" logic assumes a single buffer cache entry.
- * This seems to be left over from the ancient contiguous ramdisk
- * logic.
- */
-
sbh = CURRENT->bh;
- rbh = get_hash_table(sbh->b_dev, sbh->b_blocknr, sbh->b_size);
- if (sbh == rbh) {
- if (CURRENT->cmd == READ)
- memset(CURRENT->buffer, 1, len);
- } else if (rbh) {
- if (CURRENT->cmd == READ)
+ rbh = getblk(sbh->b_dev, sbh->b_blocknr, sbh->b_size);
+ if (CURRENT->cmd == READ) {
+ if (sbh != rbh)
memcpy(CURRENT->buffer, rbh->b_data, rbh->b_size);
- else
+ } else
+ if (sbh != rbh)
memcpy(rbh->b_data, CURRENT->buffer, rbh->b_size);
- } else { /* !rbh */
- if (CURRENT->cmd == READ)
- memset(sbh->b_data, 2, len);
- else {
- rbh = getblk(sbh->b_dev, sbh->b_blocknr, sbh->b_size);
- if (rbh)
- memcpy(rbh->b_data, CURRENT->buffer,
- rbh->b_size);
- else
- BUG(); /* No buffer, what to do here? */
- }
- }
- if (rbh) {
- mark_buffer_protected(rbh);
- brelse(rbh);
- }
+ mark_buffer_protected(rbh);
+ brelse(rbh);
end_request(1);
goto repeat;
@@ -372,6 +329,14 @@
if (DEVICE_NR(inode->i_rdev) >= NUM_RAMDISKS)
return -ENXIO;
+ /*
+ * Immunize device against invalidate_buffers() and prune_icache().
+ */
+ if (rd_inode[DEVICE_NR(inode->i_rdev)] == NULL) {
+ if((rd_inode[DEVICE_NR(inode->i_rdev)] = igrab(inode)) != NULL)
+ atomic_inc(&rd_inode[DEVICE_NR(inode->i_rdev)]->i_bdev->bd_openers);
+ }
+
MOD_INC_USE_COUNT;
return 0;
@@ -389,24 +354,31 @@
ioctl: rd_ioctl,
};
+#ifdef MODULE
/* Before freeing the module, invalidate all of the protected buffers! */
static void __exit rd_cleanup (void)
{
int i;
for (i = 0 ; i < NUM_RAMDISKS; i++) {
- struct block_device *bdev;
- bdev = bdget(kdev_t_to_nr(MKDEV(MAJOR_NR,i)));
- atomic_dec(&bdev->bd_openers);
+ if (rd_inode[i]) {
+ /* withdraw invalidate_buffers() and prune_icache() immunity */
+ atomic_dec(&rd_inode[i]->i_bdev->bd_openers);
+ /* remove stale pointer to module address space */
+ rd_inode[i]->i_bdev->bd_op = NULL;
+ iput(rd_inode[i]);
+ }
destroy_buffers(MKDEV(MAJOR_NR, i));
}
devfs_unregister (devfs_handle);
unregister_blkdev( MAJOR_NR, "ramdisk" );
blk_cleanup_queue(BLK_DEFAULT_QUEUE(MAJOR_NR));
+ hardsect_size[MAJOR_NR] = NULL;
blksize_size[MAJOR_NR] = NULL;
blk_size[MAJOR_NR] = NULL;
}
+#endif
/* This is the registration and initialization section of the RAM disk driver */
int __init rd_init (void)
@@ -441,21 +413,16 @@
S_IFBLK | S_IRUSR | S_IWUSR, 0, 0,
&fd_fops, NULL);
- hardsect_size[MAJOR_NR] = rd_hardsec; /* Size of the RAM disk blocks */
- blksize_size[MAJOR_NR] = rd_blocksizes; /* Avoid set_blocksize() check */
-
- for (i = 0; i < NUM_RAMDISKS; i++) {
- struct block_device *bdev;
+ for (i = 0; i < NUM_RAMDISKS; i++)
register_disk(NULL, MKDEV(MAJOR_NR,i), 1, &fd_fops, rd_size<<1);
- bdev = bdget(kdev_t_to_nr(MKDEV(MAJOR_NR,i)));
- atomic_inc(&bdev->bd_openers); /* avoid invalidate_buffers() */
- }
#ifdef CONFIG_BLK_DEV_INITRD
/* We ought to separate initrd operations here */
register_disk(NULL, MKDEV(MAJOR_NR,INITRD_MINOR), 1, &fd_fops, rd_size<<1);
#endif
+ hardsect_size[MAJOR_NR] = rd_hardsec; /* Size of the RAM disk blocks */
+ blksize_size[MAJOR_NR] = rd_blocksizes; /* Avoid set_blocksize() check */
blk_size[MAJOR_NR] = rd_kbsize; /* Size of the RAM disk in kB */
/* rd_size is given in kB */
@@ -468,8 +435,8 @@
#ifdef MODULE
module_init(rd_init);
-#endif
module_exit(rd_cleanup);
+#endif
/* loadable module support */
MODULE_PARM (rd_size, "1i");
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Wed Mar 15 2000 - 21:00:15 EST